π Read β 3 items
π MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments β Open platform unifying GUI agent evaluation, data collection, and online RL training across mobile and computer environments.
π Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning β TreeOPSD: tree-structured self-distillation for long-horizon agentic RL that trains without a critic model.
π Stack Overflow for Agents triples signups with ChatGPT plugin β OpenAI customer story: Stack Overflow for Agents tripled signups after shipping a ChatGPT plugin.
Both papers target RL for agents β MAI-Gym supplies the open GUI eval and training environment, TreeOPSD the long-horizon training method that drops the critic.
π MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments β Open platform unifying GUI agent evaluation, data collection, and online RL training across mobile and computer environments.
π Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning β TreeOPSD: tree-structured self-distillation for long-horizon agentic RL that trains without a critic model.
π Stack Overflow for Agents triples signups with ChatGPT plugin β OpenAI customer story: Stack Overflow for Agents tripled signups after shipping a ChatGPT plugin.
Both papers target RL for agents β MAI-Gym supplies the open GUI eval and training environment, TreeOPSD the long-horizon training method that drops the critic.
huggingface.co
Paper page - MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
Join the discussion on this paper page
π¬ Community β 1 item
π¬ Websites are closing their doors to AI agents β HN discussion of The Verge's report β 153 points, 149 comments; the top comment calls out a 'you can't scrape us but we can scrape you' mentality and small businesses whose sites 'get absolutely hammered'.
π¬ Websites are closing their doors to AI agents β HN discussion of The Verge's report β 153 points, 149 comments; the top comment calls out a 'you can't scrape us but we can scrape you' mentality and small businesses whose sites 'get absolutely hammered'.
βοΈ Ship β 3 items
π¦ Amp is now an agent in Linear β Assigning an issue to Amp starts a new Agent Session in Linear, mentioning Amp kicks one off, and the session appears in your Amp sidebar, fully viewable in Amp.
π¦ Kiro 1.4.0 β Memory Across Chat Sessions β Brings Memory to every Kiro IDE user; also makes find more reliable in long Markdown Review documents and fixes Kiro failing to start after a .deb install on Ubuntu 24.04.
π¦ Copilot model policy API for enterprise, orgs, and teams β You can now use the REST API to retrieve and update Copilot model policies for enterprises, orgs, and teams β model policy management is scriptable.
π¦ Amp is now an agent in Linear β Assigning an issue to Amp starts a new Agent Session in Linear, mentioning Amp kicks one off, and the session appears in your Amp sidebar, fully viewable in Amp.
π¦ Kiro 1.4.0 β Memory Across Chat Sessions β Brings Memory to every Kiro IDE user; also makes find more reliable in long Markdown Review documents and fixes Kiro failing to start after a .deb install on Ubuntu 24.04.
π¦ Copilot model policy API for enterprise, orgs, and teams β You can now use the REST API to retrieve and update Copilot model policies for enterprises, orgs, and teams β model policy management is scriptable.
Ampcode
Assign It To Amp
Amp is now an agent in Linear. Assign or delegate an issue to Amp, or mention @Amp, and Linear starts an Agent Session that spawns an Amp orb β with optional one-shot triage orbs for every new issue.
π Read β 4 items
π Self-Evolving Coding Agents via Harness-Policy Optimization β Self-HPO lets the coding agent improve its own harness, then trains the policy β a recursive framework that alternates harness and policy optimization for coding agents.
π Ο-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions β Benchmark exposing LLM agents' interaction fragility in long-horizon tool-calling tasks under real-world user emotions.
π Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries β Trains builder agents with reinforcement learning to reorganize document corpora into reusable hierarchical workspaces that anticipate future queries.
π New paper: Emerging trends in AI agent evaluation β Chen, Rabanser, Kapoor, and Narayanan map the methods, questions, and ecosystem of AI agent evaluation β amid growing calls for independent eval, what constitutes a 'good' one remains unclear.
Two of today's papers train agents to reshape their own working environment β Self-HPO has the coding agent improve its own harness, Librarian Agents reorganizes document corpora into workspaces that anticipate future queries.
π Self-Evolving Coding Agents via Harness-Policy Optimization β Self-HPO lets the coding agent improve its own harness, then trains the policy β a recursive framework that alternates harness and policy optimization for coding agents.
π Ο-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions β Benchmark exposing LLM agents' interaction fragility in long-horizon tool-calling tasks under real-world user emotions.
π Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries β Trains builder agents with reinforcement learning to reorganize document corpora into reusable hierarchical workspaces that anticipate future queries.
π New paper: Emerging trends in AI agent evaluation β Chen, Rabanser, Kapoor, and Narayanan map the methods, questions, and ecosystem of AI agent evaluation β amid growing calls for independent eval, what constitutes a 'good' one remains unclear.
Two of today's papers train agents to reshape their own working environment β Self-HPO has the coding agent improve its own harness, Librarian Agents reorganizes document corpora into workspaces that anticipate future queries.
huggingface.co
Paper page - Self-Evolving Coding Agents via Harness-Policy Optimization
Join the discussion on this paper page
βοΈ Ship β 2 items
π¦ GitHub Copilot models deprecated on October 19, 2026 β As of October 19, 2026, GitHub has deprecated a set of models across most Copilot experiences; the changelog lists the retired models β check anything pinned to a specific model pick.
π¦ OpenAI reportedly prepares GPT-6.1 Terra for the upcoming release β GPT-6.1 Terra has reportedly surfaced in a Microsoft pull request, and The Information reported a new Terra model is on the way β OpenAI looks set to bring back its Terra tier. Spotted, not a confirmed launch.
Model lineups moved in both directions: GitHub deprecated a set of Copilot models as of October 19, 2026, while GPT-6.1 Terra reportedly surfaced in a Microsoft pull request the following day.
π¦ GitHub Copilot models deprecated on October 19, 2026 β As of October 19, 2026, GitHub has deprecated a set of models across most Copilot experiences; the changelog lists the retired models β check anything pinned to a specific model pick.
π¦ OpenAI reportedly prepares GPT-6.1 Terra for the upcoming release β GPT-6.1 Terra has reportedly surfaced in a Microsoft pull request, and The Information reported a new Terra model is on the way β OpenAI looks set to bring back its Terra tier. Spotted, not a confirmed launch.
Model lineups moved in both directions: GitHub deprecated a set of Copilot models as of October 19, 2026, while GPT-6.1 Terra reportedly surfaced in a Microsoft pull request the following day.
The GitHub Blog
GitHub Copilot models deprecated on October 19, 2026 - GitHub Changelog
As of today, October 19, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code completions). Theseβ¦
π¬ Community β 2 items
π¬ WhatsApp's help pages hid LLM instructions: dont recommend a regulator β Instructions for WhatsApp's support LLM β including 'dont recommend a regulator' β were found embedded in the help pages; 691 points and 201 comments on HN, and a WhatsApp spokesperson told WIRED they showed up because of a bug that has now been fixed.
π¬ RuneScape Bench: MMOs for Agents [video] β Agents benchmarked inside an open-world MMO β 234 points and 80 comments on HN, with the author answering questions in-thread and the submission linking a related LostCityRS Server repo.
In both threads the party behind the system responded directly β the RuneScape Bench author answered questions in-thread, and WhatsApp's spokesperson told WIRED the hidden instructions showed up because of a bug that has now been fixed.
π¬ WhatsApp's help pages hid LLM instructions: dont recommend a regulator β Instructions for WhatsApp's support LLM β including 'dont recommend a regulator' β were found embedded in the help pages; 691 points and 201 comments on HN, and a WhatsApp spokesperson told WIRED they showed up because of a bug that has now been fixed.
π¬ RuneScape Bench: MMOs for Agents [video] β Agents benchmarked inside an open-world MMO β 234 points and 80 comments on HN, with the author answering questions in-thread and the submission linking a related LostCityRS Server repo.
In both threads the party behind the system responded directly β the RuneScape Bench author answered questions in-thread, and WhatsApp's spokesperson told WIRED the hidden instructions showed up because of a bug that has now been fixed.
π Read β 4 items
π MemT-1.0: Towards Memory-Native Evolution β Proposes a memory-native architecture that embeds continuous memory directly into foundation-model computation to improve long-context efficiency β relevant to any agent stack that leans on memory for long runs.
π Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides β arXiv
π Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents β Targets the failure mode where small language model agents get stuck in repetition loops β a distillation method that detects and penalizes stagnation during training to improve reasoning efficiency.
π CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents β Synthesizes executable training environments for long-horizon AI coworkers using workflow-guided task construction and programmatic validation β takes on the 'how do we build training environments for AI coworkers' problem.
MemT-1.0, Progress-Aware Distillation, and CoworkEnv all push on agents staying effective over long runs, at three different layers: memory architecture, training objective, and training environment.
π MemT-1.0: Towards Memory-Native Evolution β Proposes a memory-native architecture that embeds continuous memory directly into foundation-model computation to improve long-context efficiency β relevant to any agent stack that leans on memory for long runs.
π Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides β arXiv
cs.AI paper (submitted 27 Jul 2026) proposing typed MCP overrides as a mechanism for persistent human control over LLM agents β a concrete pattern for operators wiring authority and oversight into MCP-based deployments.π Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents β Targets the failure mode where small language model agents get stuck in repetition loops β a distillation method that detects and penalizes stagnation during training to improve reasoning efficiency.
π CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents β Synthesizes executable training environments for long-horizon AI coworkers using workflow-guided task construction and programmatic validation β takes on the 'how do we build training environments for AI coworkers' problem.
MemT-1.0, Progress-Aware Distillation, and CoworkEnv all push on agents staying effective over long runs, at three different layers: memory architecture, training objective, and training environment.
huggingface.co
Paper page - MemT-1.0: Towards Memory-Native Evolution
Join the discussion on this paper page
π¬ Community β 1 item
π¬ DeepSeek banned from U.S. federal civilian agencies under new DHS order β HN discussion of the new DHS order banning DeepSeek from U.S. federal civilian agencies β 549 points and 501 comments, with the Federal Register document (2026-21447) linked in the thread.
π¬ DeepSeek banned from U.S. federal civilian agencies under new DHS order β HN discussion of the new DHS order banning DeepSeek from U.S. federal civilian agencies β 549 points and 501 comments, with the Federal Register document (2026-21447) linked in the thread.
βοΈ Ship β 2 items
π¦ Gemini CLI 0.65.0 β Fixes deletion of resumed session history on quick exit (#29198) and duplicate tool response turns when resuming sessions; also enforces read-only workspace settings in untrusted folders, prevents infinite OAuth verification and retry loops, and adds support for custom OTLP headers in telemetry configuration.
π¦ Kiro 1.4.10 β Memory Setting Defaults to Off, Workflow Plan Feedback, and Resilient Powers β The Memory setting's default changes to Off, approved workflow steps follow your plan feedback, and Powers keep working when a server is unavailable; several chat fixes also land.
Both releases change default agent behavior: Gemini CLI 0.65.0 enforces read-only workspace settings in untrusted folders, and Kiro 1.4.10 changes the Memory setting default to Off.
π¦ Gemini CLI 0.65.0 β Fixes deletion of resumed session history on quick exit (#29198) and duplicate tool response turns when resuming sessions; also enforces read-only workspace settings in untrusted folders, prevents infinite OAuth verification and retry loops, and adds support for custom OTLP headers in telemetry configuration.
π¦ Kiro 1.4.10 β Memory Setting Defaults to Off, Workflow Plan Feedback, and Resilient Powers β The Memory setting's default changes to Off, approved workflow steps follow your plan feedback, and Powers keep working when a server is unavailable; several chat fixes also land.
Both releases change default agent behavior: Gemini CLI 0.65.0 enforces read-only workspace settings in untrusted folders, and Kiro 1.4.10 changes the Memory setting default to Off.
GitHub
Release Release v0.65.0 Β· google-gemini/gemini-cli
What's Changed
fix(cli): enforce read-only workspace settings in untrusted folders by @jvargassanchez-dot in #29583
fix(core): avoid duplicate tool response turns when resuming sessions by @di...
fix(cli): enforce read-only workspace settings in untrusted folders by @jvargassanchez-dot in #29583
fix(core): avoid duplicate tool response turns when resuming sessions by @di...
π Read β 2 items
π ALTK-Evolve: How learned memory tops the ΟΒ²-bench leaderboard on customer service tasks without a frontier model β IBM Research's ALTK-Evolve tops the ΟΒ²-bench leaderboard on customer service tasks with learned memory and no frontier model β relevant reading for anyone designing agent memory.
π JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs β Delegates agent tool choices to small decision models via fitted confidence thresholds, cutting cost and latency β a benchmark and deployment harness for decision models that share control with LLMs.
Both items show agent results without leaning on frontier-scale models β ALTK-Evolve tops the ΟΒ²-bench leaderboard on customer service tasks via learned memory, while JevAgentHarness cuts cost and latency by handing tool choices to small decision models.
π ALTK-Evolve: How learned memory tops the ΟΒ²-bench leaderboard on customer service tasks without a frontier model β IBM Research's ALTK-Evolve tops the ΟΒ²-bench leaderboard on customer service tasks with learned memory and no frontier model β relevant reading for anyone designing agent memory.
π JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs β Delegates agent tool choices to small decision models via fitted confidence thresholds, cutting cost and latency β a benchmark and deployment harness for decision models that share control with LLMs.
Both items show agent results without leaning on frontier-scale models β ALTK-Evolve tops the ΟΒ²-bench leaderboard on customer service tasks via learned memory, while JevAgentHarness cuts cost and latency by handing tool choices to small decision models.
huggingface.co
ALTK-Evolve: How learned memory tops the ΟΒ²-bench leaderboard on customer service tasks without a frontier model
A Blog post by IBM Research on Hugging Face
βοΈ Ship β 1 item
π¦ ACP 1.26.0 β Adds a client system prompt schema (#2361), removes next edit suggestions (#2266), and clarifies that a forked session is active (#2359) β all three changes are unstable-spec only.
π¦ ACP 1.26.0 β Adds a client system prompt schema (#2361), removes next edit suggestions (#2266), and clarifies that a forked session is active (#2359) β all three changes are unstable-spec only.
GitHub
Release Schema v1.26.0 Β· agentclientprotocol/agent-client-protocol
Added
(unstable) add client system prompt schema (#2361)
Fixed
(unstable) remove next edit suggestions (#2266)
Other
(unstable) clarify that a forked session is active (#2359)
(unstable) add client system prompt schema (#2361)
Fixed
(unstable) remove next edit suggestions (#2266)
Other
(unstable) clarify that a forked session is active (#2359)
π¬ Community β 2 items
π¬ Patch the Planet: global response coordination as agents uncover vulnerabilities β HN thread on OpenAI's Patch the Planet β coordinating global response as agents uncover vulnerabilities. 160 points / 95 comments; commenters are reacting to the released skills, including one that checks PHP versions.
π¬ Using Claude to reduce the attack surface of code β HN thread on Anthropic's piece about using Claude to reduce the attack surface of code. 116 points / 76 comments; discussion pulls in kernel-vulnerability cleanup experience.
Both kept threads sit at the same intersection: agents applied to code security β one on coordinating global response as agents uncover vulnerabilities, the other on using Claude to shrink code attack surface.
π¬ Patch the Planet: global response coordination as agents uncover vulnerabilities β HN thread on OpenAI's Patch the Planet β coordinating global response as agents uncover vulnerabilities. 160 points / 95 comments; commenters are reacting to the released skills, including one that checks PHP versions.
π¬ Using Claude to reduce the attack surface of code β HN thread on Anthropic's piece about using Claude to reduce the attack surface of code. 116 points / 76 comments; discussion pulls in kernel-vulnerability cleanup experience.
Both kept threads sit at the same intersection: agents applied to code security β one on coordinating global response as agents uncover vulnerabilities, the other on using Claude to shrink code attack surface.
π Read β 3 items
π MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records β MMEnterprise-Bench evaluates multimodal agents on reasoning over evolving workplace records with conflicts and missing evidence.
π Using thousands of AI agents to uncover new leads in biological data β OpenAI's writeup on using thousands of AI agents to uncover new leads in biological data.
π Benefit Street Partners answers questions faster with Codex β OpenAI customer story on Benefit Street Partners using Codex to answer questions faster.
π MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records β MMEnterprise-Bench evaluates multimodal agents on reasoning over evolving workplace records with conflicts and missing evidence.
π Using thousands of AI agents to uncover new leads in biological data β OpenAI's writeup on using thousands of AI agents to uncover new leads in biological data.
π Benefit Street Partners answers questions faster with Codex β OpenAI customer story on Benefit Street Partners using Codex to answer questions faster.
huggingface.co
Paper page - MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records
Join the discussion on this paper page
π¨ Watch β 2 items
π¨ Degraded performance for Claude Fable 5 β Elevated errors on requests to Claude Fable 5, under investigation since 19:49 UTC on Oct 21. Affects
π¨ OpenClaw: Browser CDP discovery could accept blocked WebSocket URLs (CVE-2026-62197) β CVE-2026-62197, medium severity: OpenClaw's browser CDP discovery could accept WebSocket URLs that were supposed to be blocked β URL restrictions in browser sessions could be bypassed.
π¨ Degraded performance for Claude Fable 5 β Elevated errors on requests to Claude Fable 5, under investigation since 19:49 UTC on Oct 21. Affects
claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork.π¨ OpenClaw: Browser CDP discovery could accept blocked WebSocket URLs (CVE-2026-62197) β CVE-2026-62197, medium severity: OpenClaw's browser CDP discovery could accept WebSocket URLs that were supposed to be blocked β URL restrictions in browser sessions could be bypassed.
Claude
Degraded performance for Claude Fable 5
Claude's Status Page - Degraded performance for Claude Fable 5.
π¬ Community β 2 items
π¬ Claude Duet β 289 points and 245 comments on HN for the Claude Duet announcement on
π¬ OpenAI models are coming to Google Cloud β 220 points and 94 comments on HN for OpenAI models coming to Google Cloud; the thread opens on the 'Microsoft remains OpenAI's primary cloud partner' disclaimer at the bottom of the announcement.
Both kept threads are HN discussions of AI-lab rollout news, and in both the opening comments target the fine print β the Claude Duet waitlist (why it isn't requestable from an already-signed-in Claude session) and the 'Microsoft remains OpenAI's primary cloud partner' disclaimer on the Google Cloud announcement.
π¬ Claude Duet β 289 points and 245 comments on HN for the Claude Duet announcement on
claude.com; the opening comment pushes on the waitlist β why signup isn't requestable from an existing signed-in Claude session.π¬ OpenAI models are coming to Google Cloud β 220 points and 94 comments on HN for OpenAI models coming to Google Cloud; the thread opens on the 'Microsoft remains OpenAI's primary cloud partner' disclaimer at the bottom of the announcement.
Both kept threads are HN discussions of AI-lab rollout news, and in both the opening comments target the fine print β the Claude Duet waitlist (why it isn't requestable from an already-signed-in Claude session) and the 'Microsoft remains OpenAI's primary cloud partner' disclaimer on the Google Cloud announcement.
βοΈ Ship β 2 items
π¦ Claude platform release notes β October 21, 2026 β API version
π¦ Pi 1.4.0 β Codemode parallel scripts and in-process subagents now queue for a free worker instead of failing (
π¦ Claude platform release notes β October 21, 2026 β API version
2026-10-21 is released, with new versions of the client SDKs and ant CLI that send it by default β Python SDK 2.0.0, TypeScript SDK 1.0.0, Go SDK 2.0.0, and ant CLI 2.0.0 among them. Those versions carry breaking changes that don't depend on the API version and are licensed under Apache 2.0 instead of MIT; user profiles also add enforcement_status and enforcement_actions fields.π¦ Pi 1.4.0 β Codemode parallel scripts and in-process subagents now queue for a free worker instead of failing (
PI_CODEMODE_MAX_WORKERS raises the limit), and compaction and branch summaries use the model's output limit so long reasoning no longer makes them fail. Also: codemode models.classify() accepts images for TypeSafe's Jev and Cloudflare's Clef classifier models, and a models.json configuration reference is generated from the schema.Claude Platform Docs
Claude Platform release notes
Updates to the Claude Platform, including the Claude API, client SDKs, and the Claude Console.
π Read β 2 items
π Claude Mods: Flipping the Harness Inside Out β Dan McAteer spent weeks using Claude Mods to mold Claude Code to his AI engineering workflow; the piece frames Mods as an early form of the 'Attention Interface' β flipping the harness paradigm from facing the machine to facing the human.
π AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics β Trains the planner model of a robot harness via supervised fine-tuning and multi-turn reinforcement learning across diverse harnesses β coding-agent post-training methods applied to agentic robotics.
Both pieces put the harness at the center: Claude Mods flips it to face the human, while AgenticRoboRL brings coding-agent post-training β supervised fine-tuning plus multi-turn RL β to planner models inside robot harnesses.
π Claude Mods: Flipping the Harness Inside Out β Dan McAteer spent weeks using Claude Mods to mold Claude Code to his AI engineering workflow; the piece frames Mods as an early form of the 'Attention Interface' β flipping the harness paradigm from facing the machine to facing the human.
π AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics β Trains the planner model of a robot harness via supervised fine-tuning and multi-turn reinforcement learning across diverse harnesses β coding-agent post-training methods applied to agentic robotics.
Both pieces put the harness at the center: Claude Mods flips it to face the human, while AgenticRoboRL brings coding-agent post-training β supervised fine-tuning plus multi-turn RL β to planner models inside robot harnesses.
Latent
Claude Mods: Flipping the Harness Inside Out
Claude Mods is an early form of the Attention Interface. It flips the agent harness paradigm from the harness facing the machine, to the harness facing the human.