https://agentstackindex.com/radar/Frontier Radar from AgentStack Indexenhttps://agentstackindex.com/radar/2026-11-13-google-antigravity-is-rolling-out-dynamic-workflows-where-the-agent/https://agentstackindex.com/radar/2026-11-13-google-antigravity-is-rolling-out-dynamic-workflows-where-the-agent/Fri, 13 Nov 2026 12:00:00 +0000Google says it is rolling out Dynamic Workflows in Google Antigravity. When you use the /run-workflow command, the agent writes a Python script that runs subagents in parallel and in sequence, with structured handoffs between them. Google says this gives deterministic control flow instead of prompt drift. You can review the script before it runs and save it to run the same flow again later. Source: AI Tracker, summarizing Google Antigravity's announcement on X.Agent platformshttps://agentstackindex.com/radar/2026-11-12-caveman-launches-caveman-code-which-runs-claude-codex-and-cursor-side/https://agentstackindex.com/radar/2026-11-12-caveman-launches-caveman-code-which-runs-claude-codex-and-cursor-side/Thu, 12 Nov 2026 12:00:00 +0000Caveman is the YC F26 company behind the token-saving Caveman prompt. It launched Caveman Code, a Mac and browser app where Claude, Codex, and Cursor work side by side on the subscriptions and keys you already pay for. A "Chief" starts chats, answers their questions, and reports back when work is done. Each pull request comes with a short review guide. Agents can also drive a desktop and browser on a cloud computer, which is in beta. In the company's own list-price runs, its harness was 21% cheaper than Codex on Terminal-Bench 4 with GPT-6.1 Sol, and both solved 90 of 192 runs. On Claude Opus 5.5 it was 14% cheaper while solving more tasks, 97 vs Claude Code's 83. The company also says it was 76% cheaper than Claude Code over day-long sessions across connected apps. It's free if you bring your own keys and logins. Source: Launch YC.Agent platformsPricing changeshttps://agentstackindex.com/radar/2026-11-12-perplexity-s-decider-v1-3-and-decider-flash-v1-3-can-now-pull-exact/https://agentstackindex.com/radar/2026-11-12-perplexity-s-decider-v1-3-and-decider-flash-v1-3-can-now-pull-exact/Thu, 12 Nov 2026 12:00:00 +0000Perplexity released Decider v1.3 and Decider Flash v1.3, the follow-ups to its earlier decision models. Instead of writing text, they return probabilities. They answer yes/no, multiple-choice, and score questions about a text or image input, which makes them a good fit for routing, classification, and checks inside agents. The new Locate feature returns exact text spans with start and end character offsets and confidence scores. It works on text only, with up to 128 Locate questions per request. The model card says v1.3 is built on Qwen3.8-27B and needs about 49 GiB of BF16 weights. Flash is distilled from it. Perplexity says both are on Hugging Face under an Apache license, through OpenRouter, and in the Perplexity Decisions API. But the Hugging Face model cards currently say the repositories are private and need an access token. MarkTechPost reports API pricing of $0.02 (v1.3) and $0.01 (Flash) per 1M input tokens, with output free. Source: Perplexity blog.ModelsPricing changeshttps://agentstackindex.com/radar/2026-11-12-replit-agent-can-now-buy-supplies-for-your-business-in-the-us-through/https://agentstackindex.com/radar/2026-11-12-replit-agent-can-now-buy-supplies-for-your-business-in-the-us-through/Thu, 12 Nov 2026 12:00:00 +0000Replit gave Replit Agent two new jobs. The first is buying things. Agent finds products, and a browser subagent (using a browser from Browserbase) fills the cart and sets up the merchant's checkout. Then it waits for you to approve the final total through Link, Stripe's payment network, with Visa supporting the payment flow. Replit says Agent never gets open-ended permission to spend. It also says shell access and general browsing are turned off during payment, and approval checks are enforced in code. Agentic Commerce is available to Core and Pro users, starting in the US, and only for one-time purchases. The second job is ads. Agent can connect or create a Google Ads account, draft ad copy, generate images, recommend daily budgets, and add conversion tracking. You review and approve each campaign before it launches across Search, YouTube, Display, Discover, and more. Source: Replit blog.Agent platformshttps://agentstackindex.com/radar/2026-11-12-edge0-open-sources-edge0-145b-a-145b-parameter-model-it-says-runs/https://agentstackindex.com/radar/2026-11-12-edge0-open-sources-edge0-145b-a-145b-parameter-model-it-says-runs/Thu, 12 Nov 2026 12:00:00 +0000On November 12, Edge0 open-sourced Edge0-145B and Edge0-120B, along with its inference framework and deployment tools. It calls Edge0-145B the first open-source on-device model with 145 billion parameters. On a compatible laptop such as a MacBook Pro, it reports 29GB peak memory and more than 80 tokens/s decoding. It also reports 1,100+ tokens/s prefill and a 128K-256K context. Edge0 says the model can plan, write and run code, debug, and check results for coding and agent work. It says there are no cloud token fees and core features work offline. The performance figures are Edge0's own. Source: Edge0 blog.Modelshttps://agentstackindex.com/radar/2026-11-12-anthropic-opens-a-private-beta-of-self-hosted-claude-managed-agents/https://agentstackindex.com/radar/2026-11-12-anthropic-opens-a-private-beta-of-self-hosted-claude-managed-agents/Thu, 12 Nov 2026 12:00:00 +0000Anthropic released a self-hosted version of Claude Managed Agents as a private beta on November 12. Companies can run the agent harness, sandboxes, credentials, and session history in their own cloud or on-premises. They also choose where model requests go: Anthropic directly, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or their own gateway. Developers build agents with the same core APIs as the hosted service. Access is by request through an Anthropic account team. Source: Anthropic.Agent platformshttps://agentstackindex.com/radar/2026-11-12-google-launches-gemini-4-hydrogen-a-fast-model-for-coding-and-agentic/https://agentstackindex.com/radar/2026-11-12-google-launches-gemini-4-hydrogen-a-fast-model-for-coding-and-agentic/Thu, 12 Nov 2026 12:00:00 +0000On November 12, Google introduced Gemini 4 Hydrogen, the fast member of the Gemini 4 family. It's built for quick work across everyday coding and agentic workflows, while Gemini 4 Carbon and Argon handle deep, long-horizon reasoning. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, and 54 at xhigh. Argon scores 50. Hydrogen has a 1M-token context window. It costs a flat $1.50 per million input tokens and $7.50 per million output tokens across that window. Starting today, it is rolling out globally across all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, and Google Search. Paid Google AI Pro and Ultra subscribers can try it in Google Antigravity. Source: Google blog.ModelsPricing changeshttps://agentstackindex.com/radar/2026-11-12-expo-launches-cloud-ios-simulators-and-android-emulators-for-ai-coding/https://agentstackindex.com/radar/2026-11-12-expo-launches-cloud-ios-simulators-and-android-emulators-for-ai-coding/Thu, 12 Nov 2026 12:00:00 +0000On November 12, Expo released Expo Simulators, cloud iOS simulators and Android emulators for coding agents such as Cursor, Claude Code, and Codex. Agents can reproduce a bug, test a fix, and verify it without a developer's machine or a Mac. Each session leaves a screen recording, screenshots, and logs on the pull request. Teammates can open the running build from a link, with no TestFlight build needed. Expo adds no AI model of its own. It is available now on Expo paid plans at $0.10 per simulator minute for iOS and Android. Starter, Production, and Enterprise include 60 minutes a month. Expo says Agent Sessions will follow, with details expected by the end of November. That's a layer for triggering agents from sources like customer feedback, with prompt-injection protection. Source: Expo press release.Agent platformsPricing changeshttps://agentstackindex.com/radar/2026-11-12-nice-launches-next-generation-cognigy-with-reusable-agent-skills-ai/https://agentstackindex.com/radar/2026-11-12-nice-launches-next-generation-cognigy-with-reusable-agent-skills-ai/Thu, 12 Nov 2026 12:00:00 +0000On November 12, NiCE released a new generation of its Cognigy platform for building and orchestrating enterprise AI agents. Agents are built from skills, tools, and personas that teams create once and reuse across agents. They can coordinate specialized sub-agents and connect to outside agents through open standards, including customers' own personal AI agents. The release adds a redesigned, AI-assisted Agentic Experience Studio with default STT, TTS, and LLM models out of the box. It also adds native access to premium models from Deepgram, ElevenLabs, and Anthropic, and an AI Assist companion across the agent lifecycle. Every step an agent takes can be traced, and voice gets native speech-to-speech support. Source: NiCE press release.Agent platformshttps://agentstackindex.com/radar/2026-11-11-google-launches-the-gemini-decisions-api-and-gemini-decision-1-which/https://agentstackindex.com/radar/2026-11-11-google-launches-the-gemini-decisions-api-and-gemini-decision-1-which/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Google DeepMind introduced the Gemini Decisions API, a new endpoint for fast, closed-set decisions. It runs on a new model, Gemini Decision 1 (gemini-decision-1). Instead of generating text, it scores candidate options in a single pass and returns probabilities and confidence scores. It's aimed at tool selection, policy checks, model routing, and context grading inside agents. One request takes a shared input of text, images, audio, video, or PDFs (up to 256K tokens). That input comes with up to 1,000 yes/no, multiple-choice, or rubric-score questions. Google says end-to-end latency is 150-250 ms. Pricing is $0.10 per 1M input tokens, with no output tokens billed, and there is a no-cost tier in Google AI Studio. Source: Google blog.ModelsPricing changeshttps://agentstackindex.com/radar/2026-11-11-zyphra-releases-zonos3-mini-an-apache-2-0-voice-cloning-tts-model-with/https://agentstackindex.com/radar/2026-11-11-zyphra-releases-zonos3-mini-an-apache-2-0-voice-cloning-tts-model-with/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Zyphra released ZONOS3-Mini, an open-source text-to-speech model for zero-shot voice cloning and fine-grained vocal control across 43 languages. It uses a sparse mixture-of-experts backbone with 100M active and 800M total parameters. Zyphra reports under 570 ms time-to-first-audio in-browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. It reports 80 ms median on a single H100. Weights are on Hugging Face under Apache 2.0, with MIT-licensed runtimes for SGLang-Omni, Apple MLX, C++/ggml, and in-browser WebGPU. That makes it worth a look for voice agents that need fast or local speech. Source: Zyphra on Hugging Face.Modelshttps://agentstackindex.com/radar/2026-11-11-gemini-api-managed-agents-can-now-run-their-tools-in-your-own-sandbox/https://agentstackindex.com/radar/2026-11-11-gemini-api-managed-agents-can-now-run-their-tools-in-your-own-sandbox/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Google added self-hosted environments to Managed Agents on the Gemini API. Google still runs the agent loop and the model. But the agent's shell and file tools can now run in compute you control instead of the default Google-hosted Linux sandbox. Your sandbox connects over outbound HTTPS using the open-source Gemini API CLI (gemini-api environments connect), with no inbound firewall ports needed. Google's docs include provider examples for Cloudflare Sandbox, Daytona, E2B, Fly.io, Modal, Vercel Sandbox, Google Cloud Run, and Agent Substrate. The runtime can also run in Docker or rootless Podman. Hosted-sandbox options such as sources, network, env, and file snapshots apply only to Google-hosted environments. Source: Gemini API docs.Agent platformshttps://agentstackindex.com/radar/2026-11-11-vapi-opens-early-access-to-agent-loop-a-harness-that-builds-scores-and/https://agentstackindex.com/radar/2026-11-11-vapi-opens-early-access-to-agent-loop-a-harness-that-builds-scores-and/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Vapi launched Agent Loop in early access. Vapi says general availability is planned for early next year. You give it a goal, instructions, knowledge, and tools, and it builds a voice agent. It scores every call against the goal, finds the calls that missed, and drafts a fix. Then it tests the fix, ships it to a slice of live traffic, checks whether the number moved, and repeats. It can adjust models, architectures, voices, configurations, and tools. The agent stays editable in the regular Vapi dashboard. Vapi says the underlying features will roll into the dashboard for existing Vapi users over the coming months. Agent Loop itself has a waitlist. No pricing was announced. Source: Vapi blog.Agent platformshttps://agentstackindex.com/radar/2026-11-11-baidu-releases-ernie-5-5-flash-preview-a-280b-moe-with-14b-active/https://agentstackindex.com/radar/2026-11-11-baidu-releases-ernie-5-5-flash-preview-a-280b-moe-with-14b-active/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Baidu released ERNIE 5.5 Flash Preview, the first post-trained model built on its ERNIE 5.5 Flash Base. It has 280B total parameters with 14B active and a 1M-token context. Baidu pitches it for coding (including long-horizon tasks and game development in Unreal Engine, Unity, and Godot), 3D modeling, knowledge work, and content creation. Every showcased piece was completed inside an agent harness. The API is available now on Baidu's Qianfan platform. Baidu says the weights will be open-sourced under Apache 2.0 soon. Baidu calls it an intermediate checkpoint, and post-training is still ongoing before the official release. No pricing was stated on the announcement page. Source: Baidu ERNIE release page.Modelshttps://agentstackindex.com/radar/2026-11-11-alibaba-opens-a-two-week-qwen4-max-preview-with-a-1m-token-context/https://agentstackindex.com/radar/2026-11-11-alibaba-opens-a-two-week-qwen4-max-preview-with-a-1m-token-context/Wed, 11 Nov 2026 12:00:00 +0000On November 11, Alibaba Cloud opened Qwen4-Max-Preview (model ID qwen4-max-preview), the next-generation flagship multimodal model in the Qwen family. It takes text, images, and video. It thinks by default, and you can adjust how much it reasons. It supports function calling, web search, built-in tools, and structured outputs. It works with the OpenAI Chat Completions, Responses, and Anthropic Messages APIs. It has a 1M-token context and up to 393,216 output tokens, and thinking and the answer share that limit. Access is through Token Plan only, with no pay-as-you-go option. The preview runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8) (11:00 AM ET on November 25), and the model is retired when it ends. Source: Alibaba Cloud Model Studio docs.Modelshttps://agentstackindex.com/radar/2026-11-10-copilotkit-releases-openbot-1-0-an-mit-licensed-self-hostable/https://agentstackindex.com/radar/2026-11-10-copilotkit-releases-openbot-1-0-an-mit-licensed-self-hostable/Tue, 10 Nov 2026 12:00:00 +0000On November 10, CopilotKit launched OpenBot 1.0, a self-hostable workspace where teams create bots for specific jobs, connect them to tools, and work alongside them. The application source is MIT-licensed. For personal use, a desktop app includes everything you need to run the workspace on a Mac or Windows PC. OpenBot for Business installs on your own Kubernetes cluster (AWS, Azure, Google Cloud, or any cluster) with one Helm chart. You bring your own models and agents, including agents built with frameworks that speak AG-UI. Per-tool permissions (allow, require approval, deny) and an audit trail control what bots can do. It relies on CopilotKit Intelligence, which is licensed separately. Source: CopilotKit blog.Agent platformshttps://agentstackindex.com/radar/2026-11-10-parcha-labs-releases-agentrun-an-open-source-platform-that-adds-self/https://agentstackindex.com/radar/2026-11-10-parcha-labs-releases-agentrun-an-open-source-platform-that-adds-self/Tue, 10 Nov 2026 12:00:00 +0000On November 10, Parcha Labs released AgentRun as an open-source alpha. It sits on top of an existing agent harness. That's pi by default, plus Claude Code, Codex, OpenCode, or any harness that speaks the Agent Client Protocol. It adds learning from past runs and durable execution that resumes where a run stopped. It also adds a Mission Control view of cost, tokens, and success rates, plus on-demand computers the agent can switch between tool calls. Agents can turn repeated work into reusable workflows that are callable as APIs. It launches with five templates: code review, GitHub release notes, KYB due diligence, inbound lead research, and a weekly competitor monitor. It supports schedules plus webhook, GitHub, and Slack triggers. You can run it locally with npx or self-host it on Render for about $27.50 a month with an OpenRouter key. You can also use the hosted console, which adds access controls, SSO, and swarms. Source: AgentRun blog.Agent platformshttps://agentstackindex.com/radar/2026-11-10-ibm-and-microsoft-launch-imagine-enterprise-a-platform-for-building/https://agentstackindex.com/radar/2026-11-10-ibm-and-microsoft-launch-imagine-enterprise-a-platform-for-building/Tue, 10 Nov 2026 12:00:00 +0000On November 10, IBM and Microsoft announced Imagine Enterprise, an agent platform aimed especially at regulated industries. IBM Consulting brings more than 6,000 prebuilt AI agents and assets built on IBM Consulting Advantage (Process Studio, Context Studio, and Advantage Core). Microsoft supplies Microsoft Foundry, GitHub Copilot, and Copilot Studio to coordinate agents. It adds Fabric, Microsoft 365, Entra, Purview, Defender, and Azure for data, identity, governance, security, and cloud. The first industry frameworks cover financial services, insurance, and retail. Manufacturing, energy, public sector, and mobility are named as possible next steps. Customers can have IBM Consulting deploy and run it, build it together with IBM Consulting, or manage it themselves. Forward-deployed engineers from both companies support qualifying engagements. No pricing was announced. Source: Microsoft Source press release.Agent platformshttps://agentstackindex.com/radar/2026-11-10-n8n-labs-releases-nod-an-open-source-personal-agent-that-runs-on-your/https://agentstackindex.com/radar/2026-11-10-n8n-labs-releases-nod-an-open-source-personal-agent-that-runs-on-your/Tue, 10 Nov 2026 12:00:00 +0000On November 10, n8n Labs, a new team inside n8n, released Nod as a research preview. It is a standalone app for Mac and Windows, and it doesn't need an n8n account. It is separate from n8n and from the agents you build in n8n. Nod works in its own local sandbox with its own browser. It also works in connected apps like Gmail, Calendar, Drive, and Slack, and on websites with no integration. It can run jobs on a schedule. Before it sends, pays, posts, or deletes anything you didn't ask for, a separate check confirms you asked. Nod is free and open source under Apache 2.0. You bring your own AI provider with an API key or your ChatGPT plan, or you use a local open-source model. n8n users can connect their instance so Nod can work with their workflows. Source: n8n blog.Agent platformshttps://agentstackindex.com/radar/2026-11-10-mirantis-agent-platform-goes-ga-a-self-hosted-way-to-run-governed-ai/https://agentstackindex.com/radar/2026-11-10-mirantis-agent-platform-goes-ga-a-self-hosted-way-to-run-governed-ai/Tue, 10 Nov 2026 12:00:00 +0000On November 10 at KubeCon + CloudNativeCon North America, Mirantis made Mirantis Agent Platform generally available. It gives teams a self-hosted, governed place to run AI agents in production on Kubernetes. It includes policy-based access controls, a sandbox for each agent, and spend caps at the organization, project, and agent levels. It also has full audit trails, a one-click network cutoff for all agents, and human approval before agents apply fixes. Mirantis pairs it with forward-deployed engineers who help build custom agents. Source: Mirantis press release.Agent platformshttps://agentstackindex.com/radar/2026-11-10-gemini-4-carbon-opens-to-all-paying-developers-at-2-10-per-1m-tokens/https://agentstackindex.com/radar/2026-11-10-gemini-4-carbon-opens-to-all-paying-developers-at-2-10-per-1m-tokens/Tue, 10 Nov 2026 12:00:00 +0000On November 10, Google opened up Gemini 4 Carbon, its strongest model for long-horizon agentic work. Cyber defenders got it two weeks earlier. It is now in the Gemini API for all paying developers in Google AI Studio and in Gemini for business. It has the same introductory price as Gemini 4 Argon, $2 per 1M input tokens and $10 per 1M output tokens. The introductory period ends on November 30, 2026, and then the price becomes $4 / $20. Google says Carbon uses more tokens than Argon to reason and verify. Google still recommends Argon for most software engineering and knowledge work. Carbon is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. Source: Google blog.ModelsPricing changeshttps://agentstackindex.com/radar/2026-11-10-apify-ai-can-now-build-and-orchestrate-actors-from-natural-language/https://agentstackindex.com/radar/2026-11-10-apify-ai-can-now-build-and-orchestrate-actors-from-natural-language/Tue, 10 Nov 2026 12:00:00 +0000On November 10, Apify updated Apify AI, the chat interface in Apify Console. It can now run code, orchestrate multiple Actors, and build new Actors from plain-language instructions. The result is a reusable Actor you can schedule, connect to other apps, or publish to Apify Store. Apify also made Server Actors generally available. They are always-ready services that work like APIs, so agents can make tool calls and get responses without starting a new run each time. Apify says weekly Actor runs passed 267 million as of October 2026. It connects to any AI agent through mcp.apify.com. Source: Apify press release on GlobeNewswire.Agent platformshttps://agentstackindex.com/radar/2026-11-10-meituan-open-sources-longcat-2-5-under-the-mit-license-with-weights-on/https://agentstackindex.com/radar/2026-11-10-meituan-open-sources-longcat-2-5-under-the-mit-license-with-weights-on/Tue, 10 Nov 2026 12:00:00 +0000On November 10, Meituan open-sourced LongCat-2.5, the model it introduced on November 1 for long-horizon agent and coding work. It is a 1.6-trillion-parameter native multimodal model with about 48 billion parameters active per token. The weights are at meituan-longcat/LongCat-2.5 on Hugging Face under the MIT license. Meituan says that license allows commercial inference deployment and further training without a separate license. Meituan says the model supports a 1M-token context. It uses W4A4 quantization, meaning 4-bit expert weights and activations with native INT4 compute. Meituan says downstream results stay comparable to the unquantized model. The model card points to SGLang-FluentLLM for NPU deployment, and the model supports thinking and non-thinking modes. Most benchmark scores on the model card are Meituan's own runs. Source: tech.meituan.com/2026/11/10/Meituan-LongCat-2.5-Open-Source.html.Modelshttps://agentstackindex.com/radar/2026-11-09-automattic-launches-spacefast-hosting-that-lets-the-ai-agent-you/https://agentstackindex.com/radar/2026-11-09-automattic-launches-spacefast-hosting-that-lets-the-ai-agent-you/Mon, 09 Nov 2026 12:00:00 +0000On November 9, Automattic launched Spacefast as a new company. It isn't an AI builder. You keep building in ChatGPT, Claude, Claude Code, Codex, Cursor, or another agent. Spacefast gives that agent build instructions, a database, hosting, and custom domains. Then it returns a private link your team can comment on, and the agent can turn those comments into fixes. It is an official plugin in Claude and ChatGPT, and other agents can connect through an API, CLI, or MCP server. You choose what the agent can access, and sensitive actions need approval. Built-in app sign-in, transactional email, and shared team knowledge are due in the next few weeks. It is free to start. Personal includes $1 a month of usage, Go is $5 a month, Pro is $15 per member per month, and Enterprise is custom. Source: Automattic press release on PR Newswire.Agent platformshttps://agentstackindex.com/radar/2026-11-09-anthropic-releases-claude-fable-5-2-and-mythos-5-2-cutting-fable/https://agentstackindex.com/radar/2026-11-09-anthropic-releases-claude-fable-5-2-and-mythos-5-2-cutting-fable/Mon, 09 Nov 2026 12:00:00 +0000On November 9, Anthropic released Claude Fable 5.2 (claude-fable-5-2) for demanding reasoning and long-horizon agentic work. It has a 1M-token context window and up to 128K output tokens. It costs $8 per 1M input tokens and $40 per 1M output tokens, down from $10 / $50 on Fable 5.1 and Mythos 5.1. Cache reads are $0.20 per 1M, Batch is half price, and Fast mode is $16 / $80. Fable 5.2 is available to all Claude API customers and on Amazon Bedrock, Google Cloud, and Microsoft Foundry. Claude Mythos 5.2 has the same capabilities and prices. It is limited to organizations verified through Anthropic programs such as the Cyber Verification Program. Source: platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2.ModelsPricing changeshttps://agentstackindex.com/radar/2026-11-09-meta-releases-muse-ember-a-0-10-0-50-per-1m-token-model-for-high/https://agentstackindex.com/radar/2026-11-09-meta-releases-muse-ember-a-0-10-0-50-per-1m-token-model-for-high/Mon, 09 Nov 2026 12:00:00 +0000On November 9, Meta released Muse Ember (muse-ember-1.0), its smallest and most efficient Muse model. It's built for the steps agents repeat all day, like tool calls, coding subtasks, routing, extraction, and long document parsing. It is available on the Meta Model API and Muse Code. It costs $0.10 per 1M input tokens, $0.004 per 1M cached input tokens, and $0.50 per 1M output tokens. It has a 1M-token context window, up to 512K output tokens, and Responses API support. Meta says it matches Claude Haiku 5.5's best score at about a third of the cost. Meta pitches it as a partner for Muse Wonder, where Wonder plans and Ember handles the volume. Source: research.meta.ai/blog/muse-ember.Modelshttps://agentstackindex.com/radar/2026-11-09-delight-ai-introduces-midas-an-agent-that-builds-tests-and-improves/https://agentstackindex.com/radar/2026-11-09-delight-ai-introduces-midas-an-agent-that-builds-tests-and-improves/Mon, 09 Nov 2026 12:00:00 +0000On November 9, Delight.ai introduced Midas, an AI agent inside Delight.ai that builds and improves the Delight Customer Agent through conversation. Teams describe what they need and supply policies, procedures, and tools. Midas proposes a plan for approval, writes the result into Actionbooks (the step-by-step instructions the agent follows), and maps each process as a flowchart. It tests proposed changes with simulated customer conversations, looks into failures, updates the instructions, and runs the tests again. Every change starts in Development and every version is kept. Nothing goes live until the plan is approved, the tests pass, and someone with approval rights says yes. Midas can also analyze live conversations and propose the next change. It is rolling out to customers in stages, and no pricing was announced. Source: delight.ai/blog/product/introducing-midas.Agent platformshttps://agentstackindex.com/radar/2026-11-09-thunderphone-v3-launches-zeus-an-agent-it-says-builds-tests-and-tunes/https://agentstackindex.com/radar/2026-11-09-thunderphone-v3-launches-zeus-an-agent-it-says-builds-tests-and-tunes/Mon, 09 Nov 2026 12:00:00 +0000On November 9, ThunderPhone launched v3 with Zeus, which it calls a fully agentic harness for voice AI. Zeus can create a voice agent from a conversation about your business, or from call-center recordings, scripts, and SOPs. It can also import an existing agent from Vapi, Retell, Bland, or ElevenLabs. It then connects VoIP and software integrations, runs text and audio call simulations, flags problems on live calls, and proposes A/B tests of fixes. You choose whether it asks before each change. v3 retires the Spark, Bolt, and Storm stacks for one architecture that Zeus tunes per use case. Zeus can optimize toward a hard price cap or toward conversion value minus call cost. Zeus is free for now. Calls use cost-plus pricing with a 17% to 25% margin, and US numbers are $1/month each. Source: thunderphone.com/blog/introducing-zeus.Agent platformsPricing changeshttps://agentstackindex.com/radar/2026-11-06-openrouter-launches-auto-code-and-a-rebuilt-auto-router-for-coding/https://agentstackindex.com/radar/2026-11-06-openrouter-launches-auto-code-and-a-rebuilt-auto-router-for-coding/Fri, 06 Nov 2026 12:00:00 +0000On November 6, OpenRouter announced next-generation Auto Router (`openrouter/auto`) and Auto Code (`openrouter/auto-code`). Auto Code is built for coding agents and harnesses. It treats each request as part of a session, picks models using coding-task pass rate, completion time, and cost per solved task, and follows edits, tests, and recurring errors across turns. OpenRouter says the new router foundation outperforms the prior Auto Router by up to 4×, depending on the benchmark, with ~3× faster routing at p90 in a live test (0.65s vs 1.96s). It also says Auto Code scores in line with Claude Opus 5.5 at 74% lower cost per task across five published benchmarks. There is no router surcharge. You pay the underlying model price. Governance, ZDR, and allow/deny lists still bound the pool. Source: openrouter.ai/blog/announcements/auto-and-auto-code-routers/.Agent platformshttps://agentstackindex.com/radar/2026-11-05-thinking-machines-lab-releases-inkling-2-small-a-594b-moe-scoring-31/https://agentstackindex.com/radar/2026-11-05-thinking-machines-lab-releases-inkling-2-small-a-594b-moe-scoring-31/Thu, 05 Nov 2026 12:00:00 +0000On November 5, Thinking Machines Lab introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters and 17B active. It supports image inputs, variable effort, and a 1M-token context window. Artificial Analysis reports a score of 31 on the Intelligence Index, which Thinking Machines says would make it the highest-scoring open-weights model developed outside China once weights ship. It is available now through Tinker for inference, sampling, and fine-tuning with closed weights. As of November 5, Thinking Machines plans an Apache 2.0 weight release the following week after external safety testing and red-teaming. Source: thinkingmachines.ai/news/inkling-2-small/ and artificialanalysis.ai/changelog.Modelshttps://agentstackindex.com/radar/2026-11-05-openai-releases-gpt-6-cyber-for-scoped-vulnerability-research-through/https://agentstackindex.com/radar/2026-11-05-openai-releases-gpt-6-cyber-for-scoped-vulnerability-research-through/Thu, 05 Nov 2026 12:00:00 +0000On November 5, OpenAI introduced GPT-6 Cyber, a cybersecurity model trained to find and validate exploitable vulnerabilities while staying inside researcher-set scope. OpenAI reports that it scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber. In OpenAI's alignment evaluation it avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and 100% with that prompt. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. Researchers used it to develop 948 privilege-escalation exploits in Linux and identify 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has released fixes for many of those. GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, with broader Daybreak Red access planned. Source: openai.com/index/introducing-gpt-6-cyber.Modelshttps://agentstackindex.com/radar/2026-11-05-aws-makes-web-answer-generally-available-on-amazon-bedrock-agentcore/https://agentstackindex.com/radar/2026-11-05-aws-makes-web-answer-generally-available-on-amazon-bedrock-agentcore/Thu, 05 Nov 2026 12:00:00 +0000On November 5, AWS announced general availability of Web Answer on Amazon Bedrock AgentCore. Web Answer is a managed tool that returns a synthesized, grounded, cited answer to a natural-language question in a single call. Agent builders can attach it without maintaining their own search, retrieve, synthesis, and citation pipeline. It builds on Web Search on AgentCore (launched in June). It plans its own searches, runs several in parallel, and decides when it has enough evidence. Builders control domain include and deny lists and publish-date filters. It attaches as a target on an AgentCore Gateway, is called over MCP, and authenticates with IAM, with what AWS describes as zero data egress. It is available in US East (N. Virginia) with per-tool-call pricing. Source: aws.amazon.com/about-aws/whats-new/2026/11/web-answer-amazon-bedrock/.Agent platformshttps://agentstackindex.com/radar/2026-11-05-mistral-studio-adds-managed-agents-workflows-and-apps-in-public/https://agentstackindex.com/radar/2026-11-05-mistral-studio-adds-managed-agents-workflows-and-apps-in-public/Thu, 05 Nov 2026 12:00:00 +0000On November 5, Mistral expanded Studio with managed agents, a plain-language workflow builder, hosted apps, managed search indexes, managed datasets, and offline and online evaluations. Most are in public preview. Studio apps inside Vibe Code Web are generally available. Managed agents are multi-step agents with sandboxes, subagents, skills, tools, and MCP connectors. Studio manages their runtime. Developers write agents in Python with the mistralai-agents SDK and test with `agents chat`. They then ship with `agents serve` (self-hosted) or `agents deploy` (Mistral-hosted, durable, resumes after crashes). A new Mistral MCP server lets an existing coding agent build and deploy to Studio with one command. Studio can run in the cloud, a VPC, or on-premises. Agent, model, and tool spans go to Studio as OpenTelemetry traces by default.Agent platformshttps://agentstackindex.com/radar/2026-11-05-aws-makes-kiro-bots-generally-available-in-kiro-crew-to-own-ongoing/https://agentstackindex.com/radar/2026-11-05-aws-makes-kiro-bots-generally-available-in-kiro-crew-to-own-ongoing/Thu, 05 Nov 2026 12:00:00 +0000On November 5, AWS introduced Kiro bots, generally available in open-source Kiro Crew. Each bot is a named, long-running agent with its own memory (Notes), tools, permissions, avatar, and ongoing chat. It's built around a responsibility such as issue triage, build health, code review, or release readiness. A bot can start worker-agent sessions in parallel, monitor and redirect them, and ask for approval when needed while continuing independent work. It can also run scheduled jobs from chat, for example overnight CI failures or Friday release notes. Developers create a bot in the Crew sidebar from the default agent or a custom agent, and can inspect, edit, or delete its Notes. Dashboards let a bot maintain a customized view of its activities.Agent platformshttps://agentstackindex.com/radar/2026-11-05-digitalocean-launches-cuttlefish-an-iphone-app-to-supervise-managed/https://agentstackindex.com/radar/2026-11-05-digitalocean-launches-cuttlefish-an-iphone-app-to-supervise-managed/Thu, 05 Nov 2026 12:00:00 +0000On November 5, DigitalOcean launched Cuttlefish, a native iPhone app in public preview for supervising Managed Agents sessions that already run in DigitalOcean-hosted microVM sandboxes. From the phone, you can follow live replies, reasoning, and tool calls. You can approve or deny tool calls with the full command, file path, or diff shown. High-risk requests need Face ID or a passcode. You can also steer, pause, resume, or stop a run, view artifacts, browse workspace files, checkpoint, fork, or roll back. Cuttlefish works across harnesses on Managed Agents, including Claude Code, Codex CLI, OpenCode, OpenHands, Hermes, LangGraph, and CrewAI. The same release announces Box Mount in private preview, which syncs a Box folder into a Managed Agents environment under Box permissions. Managed Agents remains in public preview.Agent platformshttps://agentstackindex.com/radar/2026-11-05-baseten-sandboxes-go-public-giving-each-agent-an-isolated-environment/https://agentstackindex.com/radar/2026-11-05-baseten-sandboxes-go-public-giving-each-agent-an-isolated-environment/Thu, 05 Nov 2026 12:00:00 +0000Baseten Sandboxes are now available for agentic applications. Each sandbox is an isolated environment where an agent can run commands and work with files. Examples include writing and running a user's website code, or downloading sources and collecting research results. Teams can create a sandbox per user or project and delete it when the work ends. Every sandbox starts from an image: a built-in Base Image with Python, or a custom image built from a Dockerfile or imported from a registry. Custom images are versioned per team and can include a Sandbox API server for MCP tools. Developers create and drive sandboxes with the Baseten CLI (`baseten sandbox create`, `exec`, `connect`, `delete`) or SDKs, and coding agents can use the same CLI commands. Sandboxes are free to use until November 16, with a default limit of 10 concurrent sandboxes per organization (support can raise it). Deleting a sandbox permanently removes its local data, so lasting results need external storage or a mounted sandbox volume.Agent platformshttps://agentstackindex.com/radar/2026-11-05-moonshot-ai-launches-kimi-hosted-agents-a-managed-harness-with/https://agentstackindex.com/radar/2026-11-05-moonshot-ai-launches-kimi-hosted-agents-a-managed-harness-with/Thu, 05 Nov 2026 12:00:00 +0000On November 5, Moonshot AI launched Kimi Hosted Agents, a cloud "Harness as a Service" so teams can run long-running and multi-agent work without building their own runtime. The service gives isolated sandboxes with network-access controls, stoppable sessions, exportable execution logs, and usage reports that split sandbox time, model inference, and plugin calls. It lists more than 140 skills and more than 30 plugins, plus starter templates (including eight finance templates), MCP tool connections, and an optional Kimi Inference Hook. The hook lets an enterprise safety service inspect task inputs and tool returns before each model call, for enterprise-verified customers only. Sign-up is at platform.kimi.com.Agent platformshttps://agentstackindex.com/radar/2026-11-04-ricursive-intelligence-releases-ripple-65b-preview-a-65b-apache-2-0/https://agentstackindex.com/radar/2026-11-04-ricursive-intelligence-releases-ripple-65b-preview-a-65b-apache-2-0/Wed, 04 Nov 2026 12:00:00 +0000On November 4, Ricursive Intelligence published Ripple-65B Preview, a dense 65B Transformer pretrained on 12T tokens with up to 1,048,576 tokens of context. The Hugging Face model card reports 80.52 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard against larger U.S. open-weight peers such as Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B. Weights are Apache 2.0 (BF16 instruct, NVFP4 instruct, and BF16 base). Hosted access includes OpenRouter via GMICloud at $0.20 / $0.60 per 1M input/output tokens (cached input $0.05). Source: huggingface.co/ricursiveAI/Ripple-65B-Preview.Modelshttps://agentstackindex.com/radar/2026-11-04-perplexity-launches-teammates-agents-for-long-running-work-on-max/https://agentstackindex.com/radar/2026-11-04-perplexity-launches-teammates-agents-for-long-running-work-on-max/Wed, 04 Nov 2026 12:00:00 +0000On November 4, Perplexity launched Teammates: personal agents that own long-running work inside Perplexity Computer. A teammate plans and coordinates projects, monitors connected tools, follows through when something changes, and brings completed work back for review. Users can assign different jobs to individual teammates that run in parallel. Teammates can also pass work to each other. They reuse Projects, skills, and memory already in Computer, and run in separate environments with separate browser sessions. They can be added to Slack channels so colleagues can hand them work or ask for updates. Teammates are initially available on the web as a research preview for individual Perplexity Max subscribers ($200/month). Mobile is planned, and Enterprise Pro and Enterprise Max are not included by default. Source: perplexity.ai/hub/blog/teammates-are-agents-for-complex-ongoing-work.Agent platformshttps://agentstackindex.com/radar/2026-11-04-oumi-launches-faie-to-build-evaluate-train-and-continuously-improve-ai/https://agentstackindex.com/radar/2026-11-04-oumi-launches-faie-to-build-evaluate-train-and-continuously-improve-ai/Wed, 04 Nov 2026 12:00:00 +0000On November 4, Oumi announced FAiE (Frontier AI Engineer), an agent that automates end-to-end agent development: defining objectives, tools, environments, evaluations, data curation, training, deployment, and monitoring. Teams can build agents in Oumi or connect existing orchestration through Bring Your Own Harness for evaluation and model training. The loop turns production failures into evaluation sets and training data. Beyond supervised fine-tuning, Oumi adds direct preference optimization (DPO) and reinforcement learning through GRPO in private preview. FAiE asks for approval before costly or side-effect actions such as training jobs, deployments, deletes, and outbound tool calls. Source: oumi.ai/product-launch/fall26 and docs.oumi.ai/guides/intro/faie.Agent platformshttps://agentstackindex.com/radar/2026-11-04-baseten-grid-opens-in-public-preview-routing-coding-tasks-across/https://agentstackindex.com/radar/2026-11-04-baseten-grid-opens-in-public-preview-routing-coding-tasks-across/Wed, 04 Nov 2026 12:00:00 +0000On November 4, Baseten launched Grid in public preview as an enterprise control layer for internal AI coding. Grid combines three pieces. Baseten Router picks an approved open- or closed-source model per request using expected task performance and cost, including retries, failed tool calls, and incomplete answers. A Gateway centralizes model access, team-level budgets, and company-managed credentials. Guardrails apply configurable safety policies to requests and responses at the gateway. Baseten reports initial benchmarks of up to about 70% lower cost per coding task with a 0.9-percentage-point quality tradeoff versus Opus 5.5, noting that savings vary by workload. At launch, Grid supports models from OpenAI, Anthropic, Gemini Enterprise Agent Platform, Amazon Bedrock, and Baseten. Developers can connect Claude Code, Codex, OpenCode, and Pi with one setup command, or point in-house Anthropic/OpenAI API clients at a base URL and model ID. Early access is through a waitlist.Agent platformshttps://agentstackindex.com/radar/2026-11-04-moonshot-publishes-open-weights-for-kimi-k3-1-pro-on-hugging-face-a-2/https://agentstackindex.com/radar/2026-11-04-moonshot-publishes-open-weights-for-kimi-k3-1-pro-on-hugging-face-a-2/Wed, 04 Nov 2026 12:00:00 +0000Moonshot AI published the open weights for Kimi K3.1 Pro at moonshotai/Kimi-K3.1-Pro on Hugging Face, fulfilling the open-weight date it set when the model launched in apps and API on October 23. The model card describes an open-weight, natively multimodal agentic MoE with 2.8 trillion total parameters and 104 billion activated, 93 layers (69 KDA plus 24 Gated DSA), a 1,048,576-token context, MoonViT-V2 vision (401M parameters), and MXFP4 weights with MXFP8 activations from quantization-aware training. Text and image are supported. The Kimi K3.1-Pro license allows use, modification, fine-tuning, and redistribution with conditions. Model-as-a-Service businesses above $20 million revenue in any consecutive 12 months need a separate Moonshot agreement before commercial use. Products above 100 million monthly active users or $20 million monthly revenue must display "Kimi K3.1-Pro" prominently. Those two conditions do not apply to internal use or to access through Moonshot's official products or certified inference partners.Modelshttps://agentstackindex.com/radar/2026-11-04-render-adds-sandboxes-workflows-and-shared-filesystems-for-running/https://agentstackindex.com/radar/2026-11-04-render-adds-sandboxes-workflows-and-shared-filesystems-for-running/Wed, 04 Nov 2026 12:00:00 +0000Render announced on November 4 three new pieces for running AI agents inside production apps. Render Sandboxes, in public beta, give each agent its own isolated Linux microVM that starts in under a second, suspends when idle, wakes with its files and memory intact, and bills for CPU actually used. Credential brokering, in private beta, keeps secrets out of the sandbox. Render Workflows, now generally available, runs multi-step agent tasks written in Python or TypeScript with queuing, retries, and saved step results. Tasks can run for up to 24 hours on up to 8 CPUs and 32 GB of RAM, billed per second. Shared Filesystems, in early access, let many sandboxes mount the same repository, dataset, or memory store.Agent platformshttps://agentstackindex.com/radar/2026-11-04-tencent-releases-hy4-its-first-hy-model-with-visual-understanding-and/https://agentstackindex.com/radar/2026-11-04-tencent-releases-hy4-its-first-hy-model-with-visual-understanding-and/Wed, 04 Nov 2026 12:00:00 +0000Tencent released Hy4 on November 4. It's the company's first Hy model that can use a computer, combining visual understanding, reasoning, and tool calls for long-horizon agent work. SiliconANGLE reports 770 billion total parameters with 49 billion active and a 1 million-token context window. Tencent says it's available now on Tencent Cloud TokenHub, CodeBuddy, and WorkBuddy, and through OpenRouter and OpenCode, and that it kept the same pricing as Hy4 Preview. It plans to release the weights and system cards in the coming days, with a Flash version to follow.Modelshttps://agentstackindex.com/radar/2026-11-04-cognition-says-its-swe-2-flash-coding-model-runs-at-one-fifth-the/https://agentstackindex.com/radar/2026-11-04-cognition-says-its-swe-2-flash-coding-model-runs-at-one-fifth-the/Wed, 04 Nov 2026 12:00:00 +0000Cognition introduced SWE-2 Flash on November 4. It's a 276B-parameter coding model with 12B active, built on Thinking Machines' Inkling-Small with a new mid-training stage before reinforcement learning. Cognition reports Terminal-Bench 4.0 rising from 0.9% for the base model to 20.0%, and FrontierCode 2.0 from 3.2% to 24.0%. It's available now in Devin Desktop and CLI at one-fifth the price and 3x the speed of SWE-2. The weights aren't public. Cognition says it's building a program to share them with trusted partners and researchers.Modelshttps://agentstackindex.com/radar/2026-11-04-dify-1-19-0-adds-nested-workflow-tracing-portable-app-packages-and/https://agentstackindex.com/radar/2026-11-04-dify-1-19-0-adds-nested-workflow-tracing-portable-app-packages-and/Wed, 04 Nov 2026 12:00:00 +0000Dify released version 1.19.0 on November 4. External traces (Langfuse and other providers) now include the persisted steps inside called Workflow Tools under the invoking Tool span, so nested LLM calls stay in one hierarchy. App packages move to format v2 and can export the workflows behind Workflow Tools (including transitive dependencies), so imports recreate missing tools in the destination workspace. Agents also get parallel multi-file uploads with atomic binding, inline PDF and syntax-highlighted previews, and difyctl commands to list, export, import, and publish workspace Skills. The bundled plugin daemon moves to 0.6.14-local, and Docker Compose 2.35.0 or later is required.Agent platformshttps://agentstackindex.com/radar/2026-11-03-google-antigravity-starts-rolling-out-remote-agents-that-keep-coding/https://agentstackindex.com/radar/2026-11-03-google-antigravity-starts-rolling-out-remote-agents-that-keep-coding/Tue, 03 Nov 2026 12:00:00 +0000Antigravity App v2.29.0 (November 3) starts rolling out Remote Agents to Google AI Pro and Ultra subscribers on personal Google Accounts. Remote Agents run conversations and tools in an isolated Google-managed cloud environment, so long-running tasks continue when you close the browser, sleep the laptop, or switch devices. You create a remote project with GitHub repos, Google Drive files, or uploads. You can manage sessions from the desktop app, the CLI (`agy --remote-agent`), or antigravity.google.com/remote on phone or desktop. Remote Agents are not available on the free plan, Google AI Plus, or business accounts (Workspace / Google Cloud). Source: antigravity.google/docs/changelog/ and antigravity.google/docs/remote-agents/.Agent platformshttps://agentstackindex.com/radar/2026-11-03-flower-labs-releases-endeavor-1-1-what-it-calls-a-frontier-model-you/https://agentstackindex.com/radar/2026-11-03-flower-labs-releases-endeavor-1-1-what-it-calls-a-frontier-model-you/Tue, 03 Nov 2026 12:00:00 +0000Flower Labs introduced Endeavor 1.1, its third model release this year after Lizzy and Endeavor 1.0. Flower reports gains over 1.0 on GPQA (92.0 to 94.4) and HumanEval (98.2 to 99.9). Endeavor 1.1 also adds image input. It can be deployed as a Flower-managed service, in a private cloud, or on premises, with a new option for inference on GPUs located in the EU. Customers can read the model's thinking traces and keep its outputs and traces to train their own models. Access is by request, and the benchmark comparisons are Flower's own.Modelshttps://agentstackindex.com/radar/2026-11-03-n8n-3-1-0-pre-release-adds-agent-version-history-app-building-in-the/https://agentstackindex.com/radar/2026-11-03-n8n-3-1-0-pre-release-adds-agent-version-history-app-building-in-the/Tue, 03 Nov 2026 12:00:00 +0000n8n published 3.1.0 on November 3, marked as a pre-release on GitHub. On the agent side, it adds Agent version history with a publication timeline and lets you publish a saved Agent version separately from your draft. It also adds MCP tools that read Agent session logs, improves episodic memory recall, shows recalled memories in session traces, and raises the Agent knowledge file upload limit to 25 MB. App building comes to the n8n Assistant too, with a new Apps area where you can preview and publish apps, plus an on/off switch for the Assistant. It follows the 3.0 beta from October 27, and stable 3.x is still due in early December.Agent platformshttps://agentstackindex.com/radar/2026-11-03-sycamore-build-goes-generally-available-building-and-testing-an-agent/https://agentstackindex.com/radar/2026-11-03-sycamore-build-goes-generally-available-building-and-testing-an-agent/Tue, 03 Nov 2026 12:00:00 +0000Sycamore Labs made Sycamore Build generally available on November 3. Employees describe a workflow, add context (documents, process maps, and approved access to existing systems), and review a plan. The platform then coordinates specialized AI agents to research, design, build, connect, and test the agent or application, while the employee reviews progress and gives feedback. Sycamore pitches the result as enterprise-grade agentic software that improves itself over time. Organizations and individuals can sign up at Sycamore.so.Agent platformshttps://agentstackindex.com/radar/2026-11-03-snowflake-adds-coco-and-cowork-bots-that-it-says-work-in-the/https://agentstackindex.com/radar/2026-11-03-snowflake-adds-coco-and-cowork-bots-that-it-says-work-in-the/Tue, 03 Nov 2026 12:00:00 +0000At Snowflake Expedition 2026 on November 3, Snowflake announced Bots for CoCo, its coding agent for data teams, and CoWork, its work agent for knowledge workers. You give a Bot a goal, such as preparing briefings for every customer meeting or keeping data pipelines healthy. It plans and does the work in the background with your exact access rights, and surfaces only when a decision or final approval is needed. The same launch adds scheduled CoCo Automations with audit trails, an Async Agent API for long multi-step work, Snowflake for Slack, CoCo for Excel and Sheets, and a mobile app. Snowflake also announced Cortex AI Gateway in public preview: one set of access policies for agents calling models and MCP tools.Agent platformshttps://agentstackindex.com/radar/2026-11-02-sierra-rebuilds-agent-studio-around-ghostwriter-so-investigating/https://agentstackindex.com/radar/2026-11-02-sierra-rebuilds-agent-studio-around-ghostwriter-so-investigating/Mon, 02 Nov 2026 12:00:00 +0000Sierra redesigned Agent Studio around Ghostwriter, its agent for building agents. Ghostwriter now follows you through journeys, simulations, and dashboards and sees what you're looking at. It replaces what was spread across Explorer, Sierra Assistant, and other specialized agents. Work happens in chat-style sessions that hold Ghostwriter's changes until someone reviews and merges them, and sessions can be shared or forked. Related agents can now sit in one project with shared data and channels. Changes get line-by-line diffs, Preview Chat and Simulations, plus release governance with required approvals and controlled rollouts.Agent platformshttps://agentstackindex.com/radar/2026-11-02-anthropic-s-message-threads-beta-lets-claude-agents-send-only-new/https://agentstackindex.com/radar/2026-11-02-anthropic-s-message-threads-beta-lets-claude-agents-send-only-new/Mon, 02 Nov 2026 12:00:00 +0000Anthropic launched message threads in public beta for the Claude Messages API on November 2. An agent sends the full conversation once to start a thread, then sends only the new messages plus the ID of Claude's previous response. Anthropic says median time to first token on requests of 500K to 1M tokens fell by about 35%, and it has already switched threads on for most Claude Code users. It saves time, not money: Claude still reads the whole conversation each turn, so token usage, standard prices, and rate limits are unchanged. Threads are held encrypted in memory, expire 5 minutes after the last request, and work on the Claude Platform, Claude Platform on AWS, Microsoft Foundry, and Google Cloud.Modelshttps://agentstackindex.com/radar/2026-11-01-openai-will-pool-chatgpt-and-codex-usage-into-one-token-based/https://agentstackindex.com/radar/2026-11-01-openai-will-pool-chatgpt-and-codex-usage-into-one-token-based/Sun, 01 Nov 2026 12:00:00 +0000OpenAI's help center now says that starting in December, the usage allowance on Plus, Pro, and Business plans will be shared across ChatGPT and Codex, as OpenAI brings the agentic capabilities of Work into the main ChatGPT experience. Usage will be measured in tokens, so how much of the allowance a task uses depends on the model you choose and the task. More advanced models and complex tasks use more. Instant chat won't draw from the allowance. Combined remaining usage and reset times will show in Settings. After you hit the limit, you can wait for the reset, upgrade, or buy credits. OpenAI says it expects most users won't hit resets more often than they do now, but it hasn't published allowance sizes or an exact start date. If you run heavy Codex jobs on a ChatGPT plan, they'll draw from the same pool as your chats starting in December.Pricing changeshttps://agentstackindex.com/radar/2026-11-01-meituan-releases-longcat-2-5-a-1-6-trillion-parameter-native/https://agentstackindex.com/radar/2026-11-01-meituan-releases-longcat-2-5-a-1-6-trillion-parameter-native/Sun, 01 Nov 2026 12:00:00 +0000Meituan's LongCat team introduced LongCat-2.5 on November 1. It has 1.6 trillion total parameters with about 48 billion active per token. It reads images as discrete tokens in the same stream as text, which Meituan pairs with agentic coding for visual coding, design, and 3D tasks. Training added more long-horizon trajectories and large-scale reinforcement learning on codebase development and full-project builds, including tasks rebuilt from real agent failures. In Meituan's own table, LongCat-2.5 scores 38.5 on Terminal Bench v4.0 (Claude Opus 5 is listed at 53.9), 91.9 on BrowseComp, and 83.1 on MCP Atlas. These are vendor-reported results, mostly run in the Claude Code harness. The announcement post doesn't give pricing or say whether weights will be released.Modelshttps://agentstackindex.com/radar/2026-10-31-deepseek-open-sources-dsh-superagent-a-router-and-workers-coding-agent/https://agentstackindex.com/radar/2026-10-31-deepseek-open-sources-dsh-superagent-a-router-and-workers-coding-agent/Sat, 31 Oct 2026 12:00:00 +0000DeepSeek's Harness team published dsh-superagent on October 31 under the MIT license. It's a proactive coding agent for one GitHub repository that listens to GitHub and, optionally, Feishu. A resident Router talks to people and turns allowed requests into work items. Each item runs in its own Worker session and directory, then answers in the original conversation, for example with a pull request. A web panel shows what needs a person. DeepSeek calls it a research prototype for reference only, not an official product, and says it differs from the internal version it runs day to day.Agent platformshttps://agentstackindex.com/radar/2026-10-30-microsoft-s-mage-vl-cowork-is-a-4b-computer-use-agent-that-both-clicks/https://agentstackindex.com/radar/2026-10-30-microsoft-s-mage-vl-cowork-is-a-4b-computer-use-agent-that-both-clicks/Fri, 30 Oct 2026 12:00:00 +0000Microsoft's Mage team released Mage-VL-Cowork, a 4B-parameter vision-language model that operates a desktop by clicking and typing from screenshots and, in the same task, running code and terminal commands. Microsoft reports 66.0% on OSWorld-Verified with a 50-step budget, ahead of larger open agents such as Qwen3.6-27B (63.9%) and EvoCUA-32B (56.7%), and says reinforcement learning cut model calls by 13.0% and output tokens by 25.4%. A model this small is aimed at computer-use agents that run on the device itself. The project page links the code on GitHub and the weights on Hugging Face.Modelshttps://agentstackindex.com/radar/2026-10-30-nvidia-nemoclaw-blueprints-add-nemotron-models-tuning-and-a-governed/https://agentstackindex.com/radar/2026-10-30-nvidia-nemoclaw-blueprints-add-nemotron-models-tuning-and-a-governed/Fri, 30 Oct 2026 12:00:00 +0000NVIDIA said on October 30 that NemoClaw is a set of open blueprints that build a full agent stack on top of the harness you already use: OpenClaw, Hermes Agent, or LangChain Deep Agents. Each blueprint combines Nemotron models, NeMo tools for tuning the model and harness, telemetry, and the OpenShell secure runtime. OpenShell limits what an agent can reach in files, networks, credentials, and tools. NVIDIA says LangChain tuned Deep Agents for Nemotron 3 Ultra and got near-frontier results at about 10x lower cost per run than leading closed models, with the tuned profile on GitHub. NemoClaw agents run in the cloud, in data centers, and on DGX Spark and RTX machines, and can work alongside agents like Claude, Codex, or Cursor.Agent platformsModelshttps://agentstackindex.com/radar/2026-10-29-factory-s-droid-1-0-harness-lets-a-model-run-hundreds-of-tool-calls-in/https://agentstackindex.com/radar/2026-10-29-factory-s-droid-1-0-harness-lets-a-model-run-hundreds-of-tool-calls-in/Thu, 29 Oct 2026 12:00:00 +0000Factory released Droid 1.0 on October 29. It's a model-independent coding agent harness with a new Script mode: the model writes JavaScript that runs in a sandbox and can make hundreds of tool calls in a single model request, instead of one request per call. In Factory's own Terminal-Bench 4 runs (October 23 to 28, three trials per task), Droid passed 65.2% of tasks with Claude Opus 5.5 at $4.73 per passed task, against 57.6% for Claude Code. With GPT-6.1 Sol it passed 57.6%, against 54.0% for Codex, at 12% lower cost. These are vendor-reported numbers, and Factory's methodology lists reruns and scoring adjustments. Droid 1.0 ships as a desktop app for macOS and Windows and as a terminal UI.Agent platformshttps://agentstackindex.com/radar/2026-10-29-cartesia-launches-managed-agents-a-voice-agent-platform-with-telephony/https://agentstackindex.com/radar/2026-10-29-cartesia-launches-managed-agents-a-voice-agent-platform-with-telephony/Thu, 29 Oct 2026 12:00:00 +0000Cartesia launched Managed Agents on October 29. It's an end-to-end platform for enterprise voice agents that pairs its own voice models with the agent harness. It includes telephony, knowledge bases, tool calls into your APIs and CRM during a call, and outbound calling campaigns. Cartesia says it ranks first at launch on Artificial Analysis's Speech Agent Arena and on ServiceNow's EVA-Bench. Agents can be built in 44 languages and switch between English, Spanish, Hindi, French, and Japanese mid-call, with the rest coming soon. Cartesia says the platform has already handled tens of millions of conversations in finance, retail, hospitality, and healthcare.Agent platformshttps://agentstackindex.com/radar/2026-10-29-aws-lets-ai-agents-run-step-functions-workflows-as-tools-through/https://agentstackindex.com/radar/2026-10-29-aws-lets-ai-agents-run-step-functions-workflows-as-tools-through/Thu, 29 Oct 2026 12:00:00 +0000AWS said on October 29 that you can now use your existing Step Functions workflows, both Standard and Express, as tools for AI agents through Amazon Bedrock AgentCore Gateway, which connects agents to tools over MCP. The agent decides when to run a workflow and what input to pass. Step Functions then runs each step exactly as you defined it, so a multi-step process like a refund (eligibility check, manager approval, credit) follows the same order every time. Agents can also look up workflow runs, for example to report which runs failed overnight and at which step. It's available in every AWS Region that has AgentCore Gateway, at standard Step Functions and AgentCore Gateway rates.Agent platformshttps://agentstackindex.com/radar/2026-10-29-cresta-opens-its-customer-experience-agent-platform-to-developers-with/https://agentstackindex.com/radar/2026-10-29-cresta-opens-its-customer-experience-agent-platform-to-developers-with/Thu, 29 Oct 2026 12:00:00 +0000Cresta announced on October 29 that developers can now build, test, deploy, and improve customer-facing AI agents on its platform on their own. With Cresta Conductor, which is built on the Claude Agent SDK, agents live in your own repo and CI/CD pipeline. Conductor supports the open Agent Skills standard, so skills written for coding assistants carry over. The Cresta AI Agent SDK handles custom code in standard Python. Releases can send a share of conversations to a new version, A/B test versions, and roll back, and builder and reviewer permissions control who can change each agent. Cresta also launched Cresta Academy to train and certify customer and partner engineers.Agent platformshttps://agentstackindex.com/radar/2026-10-29-mistral-releases-open-weights-for-an-improved-mistral-large-4-at-0-70/https://agentstackindex.com/radar/2026-10-29-mistral-releases-open-weights-for-an-improved-mistral-large-4-at-0-70/Thu, 29 Oct 2026 12:00:00 +0000Mistral released the weights of an improved Mistral Large 4 (Le Chonk) on October 29, three weeks after its public preview. It's a 1-trillion-parameter, natively multimodal model, and Mistral calls it the best open-weight model for coding. Mistral says it scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. It also says it scores 41% on Terminal-Bench 4.0, up 13 points from the preview. Mistral also says it scores 51% on the Artificial Analysis Cyber Index and refuses 96.6% of malicious cyber requests on average. It now has Low, Medium, and High (default) reasoning modes. You can use it in Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints, or download the weights from Hugging Face. The license is a modified MIT license. It's free to use and commercialize, except that Model-as-a-Service providers need a separate commercial license. API pricing is $0.70 per 1M input tokens, $0.07 cached input, and $2.10 output.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-28-lindy-apps-builds-internal-tools-on-your-integrations-when-you-ask-in/https://agentstackindex.com/radar/2026-10-28-lindy-apps-builds-internal-tools-on-your-integrations-when-you-ask-in/Wed, 28 Oct 2026 12:00:00 +0000Lindy said on October 28 that you can now ask it, in Slack or chat, for an internal tool your team needs, such as a deal room, a spend tracker, or an approval board. It builds a working app that saves its data and has its own link. Apps read from and write back to your integrations, so moving a deal in the app updates it in Salesforce, and each app gets its own permission for each tool it uses. You can ask for changes anytime and restore an earlier version. Everyone signs in with their Lindy account, and an app can stay private, be shared with the workspace, or be made public.Agent platformshttps://agentstackindex.com/radar/2026-10-28-voiceflow-classify-picks-intents-with-voiceflow-s-own-decision-model/https://agentstackindex.com/radar/2026-10-28-voiceflow-classify-picks-intents-with-voiceflow-s-own-decision-model/Wed, 28 Oct 2026 12:00:00 +0000Voiceflow added Voiceflow Classify on October 28. It's a new intent classification option for V3 projects that uses Voiceflow's decision model instead of a large language model. Voiceflow says it's faster, needs no training, and has no confidence threshold to tune: when no intent clearly fits, it returns no match instead of guessing. If it can't classify a message, for example because more than 254 intents are in scope, the agent falls back to RAG plus an LLM for that message. V4 projects don't use intents, so it doesn't apply to them, and existing projects keep their current setting until you switch.Agent platformshttps://agentstackindex.com/radar/2026-10-28-google-lets-android-apps-expose-functions-to-any-agent-over-mcp-with/https://agentstackindex.com/radar/2026-10-28-google-lets-android-apps-expose-functions-to-any-agent-over-mcp-with/Wed, 28 Oct 2026 12:00:00 +0000At the Android Dev Summit on October 28, Google put AppFunctions, Android's tool-calling layer, into beta so apps can share their capabilities with on-device or system agents through Android MCP. Google also added A2UI for Jetpack Compose in alpha, which lets agents reply with interactive UI built from an app's own components. And there's an alpha Agentic API in ML Kit on Gemini Nano 4 that runs functions, keeps chat history, and manages the agent loop. On the developer side, Android Studio Rabbit 2 is stable with support for Claude Agent, Codex, Antigravity, and over 40 other ACP-compliant agents, and Android CLI now has official plugins for Claude Code and Codex.Agent platformshttps://agentstackindex.com/radar/2026-10-28-google-announces-gemini-4-carbon-its-strongest-model-for-cybersecurity/https://agentstackindex.com/radar/2026-10-28-google-announces-gemini-4-carbon-its-strongest-model-for-cybersecurity/Wed, 28 Oct 2026 12:00:00 +0000Google DeepMind announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon. Google says it ranks #1 at 59 on the updated v5.0 Artificial Analysis Intelligence Index. Google also says it leads on SEC-bench Pro (88.2%) for finding vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching. It's built for long-running agentic engineering work, and Google says internal teams have seen it run for days on multi-step tasks. Access starts this week with a subset of cyber defenders, then widens to trusted partners through Google's Fairwind Program. Developers, businesses, and consumers get it later. It launches at the same introductory price as Argon: $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off, rising to $4 / $20 after the introductory period.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-28-microsoft-opens-the-copilot-harness-to-developers-in-foundry-agent/https://agentstackindex.com/radar/2026-10-28-microsoft-opens-the-copilot-harness-to-developers-in-foundry-agent/Wed, 28 Oct 2026 12:00:00 +0000Microsoft announced on October 28 that the Microsoft Copilot harness, the agent loop behind GitHub Copilot and Copilot Studio, is in public preview in Foundry Agent Service. Microsoft runs the agent loop and sandbox, and you bring instructions, reusable skills, Foundry and MCP tools with configured authentication and approvals, and a supported Foundry model. Agents can work with files and code, hand bounded tasks to built-in subagents, and keep their definitions in source control. You can set one up through the Foundry portal, azd, Foundry Toolkit for VS Code, or the GitHub Copilot app. Long-running resilience for hosted agents is also generally available. The harness runtime is free until December 1, 2026. After that, sandbox compute is $0.0994 per vCPU-hour and memory is $0.0118 per GiB-hour, with model inference and other tools billed separately.Agent platformsPricing changeshttps://agentstackindex.com/radar/2026-10-28-github-agentic-workflows-is-generally-available-running-coding-agents/https://agentstackindex.com/radar/2026-10-28-github-agentic-workflows-is-generally-available-running-coding-agents/Wed, 28 Oct 2026 12:00:00 +0000GitHub said on October 28 that GitHub Agentic Workflows, an open source project from GitHub Next, is generally available on github.com for GitHub Free, Team, and Enterprise. You describe an automation, such as issue triage, CI failure analysis, documentation updates, or test writing, in a natural-language Markdown file. It compiles to standard Actions YAML, so it reuses your existing runner groups and policy constraints. Each workflow can run on GitHub Copilot or a coding agent you bring yourself, and the default Copilot CLI engine needs a Copilot license. Agents run with read-only permissions by default inside a sandboxed container behind the Agent Workflow Firewall. Outputs go through a safe outputs check, and a threat detection job scans proposed changes before they apply. New since the public preview: self-hosted ARC runners with Docker-in-Docker, a non-experimental gh aw forecast for AI credit spend, per-workflow model providers, a model router, and micro-VM isolation.Agent platformshttps://agentstackindex.com/radar/2026-10-28-github-copilot-fusion-goes-ga-routing-each-task-across-models-with-a/https://agentstackindex.com/radar/2026-10-28-github-copilot-fusion-goes-ga-routing-each-task-across-models-with-a/Wed, 28 Oct 2026 12:00:00 +0000GitHub made Copilot Fusion generally available on October 28. Instead of picking a model, you pick Fusion and an effort level (Standard, High, or Max), and Copilot decides which models to use for each prompt: one model alone, a cheap model that escalates to a stronger one if a quality check fails, or a draft-review-revise loop between two models. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max. It did that at about half the average API cost per task, about half the tokens, and about 40% less time. You're billed for each model it uses at that model's standard rate, and for a limited time paid Copilot plans get a 30% discount on model costs with Fusion. It's on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs, and admins control which models it can use.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-27-google-deepmind-announces-genie-4-a-real-time-world-model-with-native/https://agentstackindex.com/radar/2026-10-27-google-deepmind-announces-genie-4-a-real-time-world-model-with-native/Tue, 27 Oct 2026 12:00:00 +0000Google DeepMind announced Genie 4 on October 27. It's a world model for real-time interactive simulation that generates audio natively, including ambient sound, music, and interactive dialogue. You can change a running simulation with natural language, such as altering the weather, moving the camera, or adding a person or object. Gemini 3.8 Flash acts as a guiding agent in the loop, breaking goals into steps and checking incoming frames to verify each step. DeepMind also showed Gemini 4 Argon working with Genie 4 to simulate a humanoid robot playing checkers, and it pitches Genie 4 for uses including training environments for robots and agents. For now it's a research preview for trusted testers. It's coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. DeepMind says it's still early and experimental, with details that can drift in longer sessions, and all its video and audio carries SynthID watermarks.Modelshttps://agentstackindex.com/radar/2026-10-27-microsoft-power-apps-adds-an-agentic-coauthor-inside-canvas-studio-in/https://agentstackindex.com/radar/2026-10-27-microsoft-power-apps-adds-an-agentic-coauthor-inside-canvas-studio-in/Tue, 27 Oct 2026 12:00:00 +0000Microsoft said on October 27 that agentic authoring in Power Apps canvas studio is now in preview. The authoring agent sits beside the canvas and works with existing apps, Dataverse tables, connectors, and governance. You describe what you want, it proposes a plan, and after you agree or make changes it does the work, such as replicating a change across every screen of an app. The result is still a regular canvas app, so you can choose when to spend tokens or edit by hand. An admin turns it on with a tenant setting in the Power Platform Admin Center, and makers sign in with a GitHub.com account. It follows the canvas authoring agent plugin, which became generally available in September for building canvas apps from coding agents like GitHub Copilot CLI and Claude Code.Agent platformshttps://agentstackindex.com/radar/2026-10-27-ibm-watsonx-orchestrate-can-now-find-sandbox-and-evaluate-agents-built/https://agentstackindex.com/radar/2026-10-27-ibm-watsonx-orchestrate-can-now-find-sandbox-and-evaluate-agents-built/Tue, 27 Oct 2026 12:00:00 +0000At TechXchange 2026 on October 27, IBM said watsonx Orchestrate can connect to Amazon Bedrock, Microsoft Foundry, and Google Gemini to discover agents built there, along with models and MCP servers (tech preview). It can also run agents in isolated sandboxes, with a kill switch to shut any agent down (tech preview). The evaluation tools are generally available. They test the quality, accuracy, and compliance of agents, models, and MCP servers before deployment, monitor them in production, and track each agent's model and token use. IBM also made Agent Identity and HashiCorp Vault 2.2 generally available. Together they give each agent a governed identity and short-lived, least-privilege credentials that can be revoked in real time.Agent platformshttps://agentstackindex.com/radar/2026-10-27-liquid-ai-releases-lfm3-six-open-weight-2b-class-models-for-text/https://agentstackindex.com/radar/2026-10-27-liquid-ai-releases-lfm3-six-open-weight-2b-class-models-for-text/Tue, 27 Oct 2026 12:00:00 +0000Liquid AI released LFM3 on October 27. It's a family of six open-weight models of about 2B parameters, all built on one LFM3-1.7B-Base backbone: four cover text, vision, audio, and omni inputs, and two are decision models. Liquid AI says LFM3-1.7B outperforms its earlier LFM2.5-2.6B with 35% fewer parameters. The text model has a 256K context window and fits in about 1 GB at 4-bit precision, and every model can be fine-tuned on a single GPU. You can get the models on Hugging Face, the Liquid AI Playground, and the Liquid AI API.Modelshttps://agentstackindex.com/radar/2026-10-27-n8n-3-0-beta-arrives-with-breaking-changes-docker-only-self-hosting/https://agentstackindex.com/radar/2026-10-27-n8n-3-0-beta-arrives-with-breaking-changes-docker-only-self-hosting/Tue, 27 Oct 2026 12:00:00 +0000n8n released the 3.0.0 beta on October 27, with the stable release due in early December. It adds no new features. It removes legacy nodes (Function, Item Lists, Cron, Interval, the legacy OpenAI node, older AI Agent modes), tightens security defaults, cuts the default Code node timeout from 5 minutes to 1 minute, and runs manual executions on workers in queue mode. Self-hosted n8n now needs Docker, since npm and npx installs are no longer supported. Instances on 2.39 or later have a Migration Report under Settings that flags affected workflows and fixes many automatically. Upgrading is one way. 2.x gets security and critical fixes for 12 months after the stable release, and 1.x support ends December 31, 2026.Agent platformshttps://agentstackindex.com/radar/2026-10-27-salesforce-adds-agent-evals-a-kill-switch-and-hard-spending-limits-to/https://agentstackindex.com/radar/2026-10-27-salesforce-adds-agent-evals-a-kill-switch-and-hard-spending-limits-to/Tue, 27 Oct 2026 12:00:00 +0000Salesforce expanded Agent Fabric, its control plane for governing agents across platforms, with Agent Evals and API Evals that will continuously check agent quality and cost, no matter which platform built the agent. Agent Kill Switch, open policies for federated API gateways, and a FinOps dashboard are generally available now. The dashboard comes with Model Wallet and cost optimization policies that can cap usage at budget. Agent Evals and API Evals enter beta in November 2026 and reach general availability in January 2027, and a Custom Agent Scanner is in beta today.Agent platformshttps://agentstackindex.com/radar/2026-10-26-entire-launches-a-git-forge-built-for-coding-agents-and-humans/https://agentstackindex.com/radar/2026-10-26-entire-launches-a-git-forge-built-for-coding-agents-and-humans/Mon, 26 Oct 2026 12:00:00 +0000Entire, the company led by former GitHub CEO Thomas Dohmke, has opened its Git forge to every developer. Instead of pull requests it uses Trails. Custom evals are scored by agents on every push, so coding agents can fix findings in a loop and a developer can merge once all evals pass. It also adds Continuous Memory, which syncs agent sessions into the repository so agents can search earlier work, and Entire says its codebase graph saves up to 71% on tokens. Entire says pushes are up to 89x faster than GitHub, and it offers private repos hosted in the US, EU, India, or Australia.Agent platformshttps://agentstackindex.com/radar/2026-10-26-openai-releases-gpt-6-1-terra-at-1-5-per-1m-tokens-built-for-computer/https://agentstackindex.com/radar/2026-10-26-openai-releases-gpt-6-1-terra-at-1-5-per-1m-tokens-built-for-computer/Mon, 26 Oct 2026 12:00:00 +0000GPT-6.1 Terra is OpenAI's new fast, low-cost model for visual reasoning, computer use, coding, and everyday work. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. On OpenAI's own Agents' Last Exam benchmark it scores 15 points higher than GPT-5.6 Terra at medium effort. On Terminal-Bench 4.0 it solves 50.3% of tasks at max effort, up from GPT-5.6 Terra's published 21.5%. These results are self-reported. It's available in ChatGPT Work, Codex, and the API as gpt-6.1-terra, and it launched the same day on Amazon Bedrock (with up to 1M input tokens) and in GitHub Copilot.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-26-oracle-makes-ai-database-agent-engine-generally-available-running/https://agentstackindex.com/radar/2026-10-26-oracle-makes-ai-database-agent-engine-generally-available-running/Mon, 26 Oct 2026 12:00:00 +0000Oracle AI Database Agent Engine is now generally available in Oracle AI Database 26ai. It runs agent workflows inside the database, next to the data they use. It takes agents built with Open Agent Specification and supports ReAct, sequential, parallel, orchestrator, router, and loop patterns, plus human approval gates. SQL and PL/SQL handle the deterministic steps, and language models are called only at the points the workflow defines, so teams don't need a separate orchestration layer kept in sync with their data.Agent platformshttps://agentstackindex.com/radar/2026-10-24-voiceflow-s-atlas-assistant-can-now-set-up-a-phone-number-for-your/https://agentstackindex.com/radar/2026-10-24-voiceflow-s-atlas-assistant-can-now-set-up-a-phone-number-for-your/Sat, 24 Oct 2026 12:00:00 +0000Voiceflow's in-app assistant, Atlas, can now add a phone number to an agent from a Phone number card in its chat. You can ask it to set one up, or it offers one after it builds or publishes a voice agent that has no number. Creating a new number or adding an existing one opens the same forms you'd see on the Phone numbers page. If Atlas moves, disconnects, or deletes a number that answers for an agent, an approval card names that agent and waits for you, even in Auto mode.Agent platformshttps://agentstackindex.com/radar/2026-10-24-kanzhun-s-nanbeige-team-releases-nanbeige4-3-33b-an-apache-2-0-model/https://agentstackindex.com/radar/2026-10-24-kanzhun-s-nanbeige-team-releases-nanbeige4-3-33b-an-apache-2-0-model/Sat, 24 Oct 2026 12:00:00 +0000Nanbeige4.3-33B is a dense 33B model with a 524,288-token context window. It's on Hugging Face under Apache 2.0, so you can use the weights commercially. The team pitches it for coding agents, tool use, and office tasks, and recommends its preserve_thinking option, which keeps earlier reasoning across turns, for multi-turn tool use. In its own comparison it scores 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE, ahead of Qwen3.8-27B, though its SWE-Bench Pro score of 60.3 trails Qwen's 61.7. These results are self-reported and not independently verified. A distilled 3B version is planned.Modelshttps://agentstackindex.com/radar/2026-10-23-gumloop-13-0-revamps-its-agent-harness-and-adds-org-wide-agent/https://agentstackindex.com/radar/2026-10-23-gumloop-13-0-revamps-its-agent-harness-and-adds-org-wide-agent/Fri, 23 Oct 2026 12:00:00 +0000Gumloop 13.0 (Sointula) rebuilds the harness its agents run on, so they can take on longer, more complex work and message you along the way in chat, Slack, Teams, iMessage, and email. A new Evaluation Builder turns a plain description of what to check into a single evaluation that grades several agents across the organization. Admins can now hand a member's agents, workflows, skills, triggers, and knowledge sources to a teammate in one step. Artifact analytics show who viewed an artifact, how long they stayed, and what they did. The release also adds MiniMax M3.1 Flash as a model, a connector for starting and checking Codex coding tasks, and lower costs for Claude agents that run every few minutes.Agent platformshttps://agentstackindex.com/radar/2026-10-23-dify-1-18-0-lets-agents-pick-the-next-workflow-step-and-packs-apps/https://agentstackindex.com/radar/2026-10-23-dify-1-18-0-lets-agents-pick-the-next-workflow-step-and-packs-apps/Fri, 23 Oct 2026 12:00:00 +0000Dify 1.18.0 lets an Agent node choose which workflow branch runs next: turn on Output Routes, give each route a condition, and connect it to the next step. Apps can now move between workspaces as a single .ifpkg package that carries their inline agents, skills, and files. You can also open a Workflow Tool call in the run history and see each step of the child workflow. An opt-in Agent Knowledge Base is in beta, and the new difyctl CLI can create, import, export, and publish apps. If you self-host, read the upgrade guide first. The agent components have to be upgraded together, and source installs now need Python 3.13.Agent platformshttps://agentstackindex.com/radar/2026-10-23-moonshot-releases-kimi-k3-1-pro-with-open-weights-due-by-nov-4/https://agentstackindex.com/radar/2026-10-23-moonshot-releases-kimi-k3-1-pro-with-open-weights-due-by-nov-4/Fri, 23 Oct 2026 12:00:00 +0000Kimi K3.1 Pro upgrades Moonshot's 3T-class Kimi K3 and has a 1M-token context window. It's live now in Kimi's apps and API at $2 per million input tokens and $10 per million output tokens. A Highspeed version runs 2 to 3 times faster for $4 and $20. Moonshot says it scored 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4%, but that's the company's own number. The API accepts OpenAI- and Anthropic-style requests, so you can put it behind Claude Code or Codex. Moonshot also says it stripped hacking-related training data out before the open release.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-23-minimax-publishes-open-weights-for-minimax-m3-1-flash-a-428b-moe-agent/https://agentstackindex.com/radar/2026-10-23-minimax-publishes-open-weights-for-minimax-m3-1-flash-a-428b-moe-agent/Fri, 23 Oct 2026 12:00:00 +0000You can now download MiniMax-M3.1-Flash on Hugging Face. It has about 428B total parameters with 23B active per token, takes text, image, and video input, and has a context window of up to 1M tokens. MiniMax pitches it for coding, agentic workflows, and computer use, and its own model card lists 73.7% on DeepSWE v1.1 and 61.7 on OSWorld 2.0. vLLM had serving support on day 0. If you run agents on self-hosted models, it's a big new option worth testing. Just keep in mind that MiniMax ran these benchmarks itself.Modelshttps://agentstackindex.com/radar/2026-10-22-agentic-ai-foundation-adds-agent-skills-open-responses-openenv-and/https://agentstackindex.com/radar/2026-10-22-agentic-ai-foundation-adds-agent-skills-open-responses-openenv-and/Thu, 22 Oct 2026 12:00:00 +0000The Linux Foundation's Agentic AI Foundation now hosts four more open projects. Agent Skills from Anthropic is a format for packaging instructions, scripts, and resources that agents load on demand. Open Responses from OpenAI is a multi-provider spec based on the Responses API. OpenEnv from Hugging Face covers reinforcement-learning environments. OpenSandbox from Alibaba is a sandbox platform with Docker and Kubernetes runtimes. They join A2A, Agent Router, agentgateway, AGENTS.md, goose, and MCP, bringing the foundation to 10 projects. For builders, that means the skills and model interfaces you write today have a better chance of working across vendors.Agent platformshttps://agentstackindex.com/radar/2026-10-22-google-cloud-announces-agent-compute-engine-and-wider-agent-governance/https://agentstackindex.com/radar/2026-10-22-google-cloud-announces-agent-compute-engine-and-wider-agent-governance/Thu, 22 Oct 2026 12:00:00 +0000Agent Compute Engine, announced at AGNTCon + MCPCon, is a compute service built for running agents in hardware-isolated sandboxed runtimes. Google says it already powers Gemini as an agent for work tasks. It arrives in preview in the coming months. Google is also extending the agent governance in Gemini for business to cover agents wherever they run. That matters if your team needs central control over agents built on several platforms.Agent platformshttps://agentstackindex.com/radar/2026-10-22-thoughtspot-launches-agentspot-a-free-plan-builder-for-agents/https://agentstackindex.com/radar/2026-10-22-thoughtspot-launches-agentspot-a-free-plan-builder-for-agents/Thu, 22 Oct 2026 12:00:00 +0000Anyone on a team can describe a job in plain language, and AgentSpot turns it into an agent, a scheduled or triggered workflow, or an interactive app. There's no charge for tokens or seats. You pay per agent conversation, workflow run, or app session, and building and testing are free. The free plan has no time limit, needs no credit card, and includes unlimited agents and users with a weekly usage allowance. Heavier use moves to the paid Fleet plan. It's a new no-code option for teams whose agents mostly work with company data.Agent platformsPricing changeshttps://agentstackindex.com/radar/2026-10-21-openai-models-come-to-google-cloud-s-model-garden/https://agentstackindex.com/radar/2026-10-21-openai-models-come-to-google-cloud-s-model-garden/Wed, 21 Oct 2026 12:00:00 +0000GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are in private preview on Google Cloud, with Codex agentic workflows, and general availability is planned in the coming weeks. You pay through Google Cloud Marketplace, and the spend can draw down existing GCP commitments. OpenAI models are also coming to Google Antigravity and the Gemini Agent. If your agents already run on Google Cloud, you'll soon be able to switch models without signing a separate vendor contract.ModelsAgent platformshttps://agentstackindex.com/radar/2026-10-21-anthropic-opens-a-research-preview-of-claude-duet-a-proactive-agent/https://agentstackindex.com/radar/2026-10-21-anthropic-opens-a-research-preview-of-claude-duet-a-proactive-agent/Wed, 21 Oct 2026 12:00:00 +0000Claude Duet watches your connected apps, surfaces what needs your attention, keeps working when you step away, and checks with you before it acts. It's on iOS, Android, and claude.ai, comes with Pro and Max plans, and is rolling out by waitlist, starting with US subscribers. It's built for professionals and small business owners, which puts it in the same lane as ready-made assistants like Lindy.Agent platformshttps://agentstackindex.com/radar/2026-10-21-glean-adds-headless-agent-building-for-claude-code-cursor-and-codex/https://agentstackindex.com/radar/2026-10-21-glean-adds-headless-agent-building-for-claude-code-cursor-and-codex/Wed, 21 Oct 2026 12:00:00 +0000Developers can now build, test, and publish Glean agents as code right from their coding tools, and the finished agents run on Glean's managed platform. It's generally available now. Worth a look if you'd rather keep agents versioned in a repo than build them in a visual editor.Agent platformshttps://agentstackindex.com/radar/2026-10-21-nvidia-adds-open-nemotron-3-5-retrieval-models-for-embedding-reranking/https://agentstackindex.com/radar/2026-10-21-nvidia-adds-open-nemotron-3-5-retrieval-models-for-embedding-reranking/Wed, 21 Oct 2026 12:00:00 +0000The new Nemotron Retriever models are open and commercially available. NVIDIA's pitch is that better retrieval gets agents to grounded answers with fewer repeated searches and fewer wasted tokens. Worth testing if your agents lean on RAG.Modelshttps://agentstackindex.com/radar/2026-10-20-zapier-mcp-lets-claude-chatgpt-and-cursor-share-one-server-with-one/https://agentstackindex.com/radar/2026-10-20-zapier-mcp-lets-claude-chatgpt-and-cursor-share-one-server-with-one/Tue, 20 Oct 2026 12:00:00 +0000Zapier said on October 20 that you can now connect several AI tools, such as Claude, ChatGPT, and Cursor, to one Zapier MCP server. Every client on it shares the same apps, actions, skills, and mode, and every tool call lands in one history. Mode belongs to the server. In managed mode you choose exactly which actions all the clients can run, and you can lock fields, like the Slack channel a message posts to. Skills only work in agentic mode. A client can connect to only one server at a time, and adding it to a shared server takes it off its old one. Multi-client servers are available now in Zapier MCP, which is included in every Zapier plan, and you need Owner or Editor access to add clients.Agent platformshttps://agentstackindex.com/radar/2026-10-20-mistral-releases-voxtral-3-speech-models-and-a-voice-agent-sdk/https://agentstackindex.com/radar/2026-10-20-mistral-releases-voxtral-3-speech-models-and-a-voice-agent-sdk/Tue, 20 Oct 2026 12:00:00 +0000Built on LiveKit, the SDK runs transcription, reasoning with tool calls, and speech in one streaming pipeline, and Mistral says each turn completes in under a second. Transcription covers 50+ languages, and the voice agent API starts at $0.04/min with Mistral Small 4. If you're building phone or support agents, note that it runs on Mistral's EU, US, and global endpoints or self-hosted.ModelsAgent platformshttps://agentstackindex.com/radar/2026-10-20-make-announces-makeos-and-upgrades-maia-at-waves-26/https://agentstackindex.com/radar/2026-10-20-make-announces-makeos-and-upgrades-maia-at-waves-26/Tue, 20 Oct 2026 12:00:00 +0000MakeOS is a new product for running AI agents, where every agent, action, and cost is visible, permissioned, and reversible. Maia, Make's AI scenario builder, gets plan mode, memory, image input, and error fixing. Subagent orchestration is out now, while human-in-the-loop approvals and data pinning are marked coming soon.Agent platformshttps://agentstackindex.com/radar/2026-10-19-meta-releases-muse-glimmer-30b-an-open-on-device-agent-model-under/https://agentstackindex.com/radar/2026-10-19-meta-releases-muse-glimmer-30b-an-open-on-device-agent-model-under/Mon, 19 Oct 2026 12:00:00 +0000It handles tool use, multi-step planning, and failure recovery locally, takes text and image input, and is quantized to about 4-bit so it runs on a single GPU. Meta says it beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks. Worth a look if you want your agent runs to stay on your own hardware.Modelshttps://agentstackindex.com/radar/2026-10-17-replit-makes-openai-the-default-model-for-replit-agent/https://agentstackindex.com/radar/2026-10-17-replit-makes-openai-the-default-model-for-replit-agent/Sat, 17 Oct 2026 12:00:00 +0000CEO Amjad Masad pointed to cost, the partnership, and performance, and Claude models are still selectable. If your build depends on how Claude behaves, pin the model rather than trusting the default.Agent platformsModelshttps://agentstackindex.com/radar/2026-10-16-reflection-releases-beam-an-open-weight-501b-moe-model-23b-active/https://agentstackindex.com/radar/2026-10-16-reflection-releases-beam-an-open-weight-501b-moe-model-23b-active/Fri, 16 Oct 2026 12:00:00 +0000It's built for coding and long-horizon agent work, with weights on Hugging Face and 40+ launch partners. If you self-host your agent backend, this is a permissively licensed option worth a look.Modelshttps://agentstackindex.com/radar/2026-10-15-meta-releases-muse-spark-open-816b-moe-42b-active-1m-context/https://agentstackindex.com/radar/2026-10-15-meta-releases-muse-spark-open-816b-moe-42b-active-1m-context/Thu, 15 Oct 2026 12:00:00 +0000Open-weight model built for coding and long-horizon agent work; available via OpenRouter, Together, Fireworks, Groq and others — a new self-host/API option for agent backends.Modelshttps://agentstackindex.com/radar/2026-10-15-nvidia-releases-nemotron-3-5-super-vl-and-nemotron-3-5-bolt-74b-a9b/https://agentstackindex.com/radar/2026-10-15-nvidia-releases-nemotron-3-5-super-vl-and-nemotron-3-5-bolt-74b-a9b/Thu, 15 Oct 2026 12:00:00 +0000New open-weight NVIDIA models to watch for cheap self-hosted agent backends.Modelshttps://agentstackindex.com/radar/2026-10-13-gemini-4-argon-ga-2-10-per-1m-tokens-through-nov-30-then-4-20/https://agentstackindex.com/radar/2026-10-13-gemini-4-argon-ga-2-10-per-1m-tokens-through-nov-30-then-4-20/Tue, 13 Oct 2026 12:00:00 +0000Google’s new frontier model for long-horizon agent tasks; intro pricing ends Dec 1 — watch agent builder model pickers and Gateway credit burn.ModelsPricing changeshttps://agentstackindex.com/radar/2026-10-13-n8n-adds-gemini-4-argon-for-agents-assistant-and-mcp-workflows/https://agentstackindex.com/radar/2026-10-13-n8n-adds-gemini-4-argon-for-agents-assistant-and-mcp-workflows/Tue, 13 Oct 2026 12:00:00 +0000One of the first builder platforms with Argon at launch; Cloud users can run it via Gateway credits without a separate Google key.Agent platformsModelshttps://agentstackindex.com/radar/2026-10-06-zapier-folds-zapier-agents-into-ai-by-zapier-an-agent-step-inside-any/https://agentstackindex.com/radar/2026-10-06-zapier-folds-zapier-agents-into-ai-by-zapier-an-agent-step-inside-any/Tue, 06 Oct 2026 12:00:00 +0000An agent's instructions, tools, and triggers now live in one AI step in the Zap editor. That means an agent can sit next to normal Zap steps, filters, and paths, and its runs show up in Zap History. You can require approval per tool instead of per agent, and you can pick a Standard, Advanced, or Premium model tier or name a specific model. AI steps bill as tasks. Default models are free, other tiers cost 1, 3, or 5 tasks per step, and a run that goes past 75 tasks pauses for a human. Existing Zapier Agents users will see a migration banner on eligible agents.Agent platformsPricing changeshttps://agentstackindex.com/radar/oct-2026-anthropic-releases-claude-haiku-5-5/https://agentstackindex.com/radar/oct-2026-anthropic-releases-claude-haiku-5-5/A new small, fast Claude tier — the usual pick for high-volume agent steps.Modelshttps://agentstackindex.com/radar/oct-2026-claude-opus-5-5-takes-1-on-the-artificial-analysis-intelligence-index/https://agentstackindex.com/radar/oct-2026-claude-opus-5-5-takes-1-on-the-artificial-analysis-intelligence-index/Top model got cheaper, with a larger cache-hit discount — good news for long agent loops.ModelsPricing changeshttps://agentstackindex.com/radar/oct-2026-gpt-6-sol-and-luna-halve-cost-vs-gpt-5-6/https://agentstackindex.com/radar/oct-2026-gpt-6-sol-and-luna-halve-cost-vs-gpt-5-6/Same index scores at about half the cost — check whether your agent builder lets you switch.ModelsPricing changeshttps://agentstackindex.com/radar/oct-2026-mistral-releases-mistral-large-4/https://agentstackindex.com/radar/oct-2026-mistral-releases-mistral-large-4/Artificial Analysis calls it the most intelligent model outside the US and China — an EU option for compliance-minded teams.Modelshttps://agentstackindex.com/radar/2026-gpt-6-astra-launched-at-10-50-per-1m-tokens/https://agentstackindex.com/radar/2026-gpt-6-astra-launched-at-10-50-per-1m-tokens/Ties the leaders on Artificial Analysis indices at lower cost than Claude Fable 5.1.Modelshttps://agentstackindex.com/radar/2026-gemini-3-8-flash-ga-0-75-3-75-per-1m-tokens-until-dec-31-2026-then-1/https://agentstackindex.com/radar/2026-gemini-3-8-flash-ga-0-75-3-75-per-1m-tokens-until-dec-31-2026-then-1/Launch pricing ends Jan 1 — agents running on Flash will roughly double in model cost.ModelsPricing changes