📚 Read — 3 items

📚 From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation — Proposes agentic language world models — interactive environment simulators built from traces, letting agents run in simulated worlds; relevant to anyone building agent training or eval environments.

📚 LegalOn halves Codex costs while maintaining development speed — OpenAI customer story on Codex in production: LegalOn halves costs while maintaining development speed — a concrete cost-outcome data point for coding-agent deployments.

📚 RoboQuest: Generalist Physical Agents that Search, Inspect and Test — A benchmark for goal-directed embodied exploration — generalist physical agents that search, inspect and test. Physical-agent eval rather than coding-agent, but squarely agent-benchmark research.

Two of today's three reads approach agent evaluation from opposite ends — RoboQuest defines a benchmark for goal-directed embodied exploration, while the agentic-world-models paper works on the environments themselves, simulating them from traces.

⚙️ Ship — 3 items

📦 NanoClaw 2026.10.0 — First calendar-versioned release; /update-nanoclaw now installs published releases instead of the tip of main. Breaking: OneCLI installs must upgrade the OneCLI gateway to exactly 1.42.0 first — it closes a credential-injection bypass, and /update-nanoclaw never upgrades a running gateway.

📦 Kiro — Memory Across Chat Sessions — Kiro changelog — memory across chat sessions, so context carries between sessions.

📦 Anthropic: 2026 usage policy update — Anthropic published its annual Usage Policy update on Oct 8, 2026; the post summarizes the changes.

💬 Community — 1 item

💬 Show HN: Let your AI agents paint big arrows, boxes and text on your screen — Show HN: a GitHub project that lets AI agents draw arrows, boxes, and text overlays directly on your screen — 307 points and 130 comments in its first 7 hours on HN.

📚 Read — 3 items

📚 SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces — Self-distills spatial intelligence from verified coding-agent traces; the paper opens from the premise of coding agents' widespread adoption.

📚 How to build great out-of-the-box user experiences with Managed Deep Agents — LangChain on building out-of-the-box user experiences with Managed Deep Agents — the post's URL points to a Slack SDK for the deepagents framework.

📚 SuperNav: An Agentic Navigation System for Any Task in Any Scene — Paper presents SuperNav, described as an agentic navigation system for diverse tasks and scenes.

💬 Community — 1 item

💬 South Korea says AI agents appear to have been used to hack the country's banks — South Korea says AI agents appear to have been used to hack the country's banks, per Reuters; the HN discussion sits at 100 points and 32 comments.

📚 Read — 3 items

📚 Anthropic research: Investigating unintended model actions — Describes examples of unintended model actions observed during evaluations and internal use of Claude — part of Anthropic's effort to publish more frequent standalone reports on model behavior and alignment.

📚 Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments? — Introduces Learn2Play Bench, a benchmark measuring how well LLM agents learn from experience in unfamiliar environments.

📚 MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement — Technical report of the MiMo V2.6 series — scaling reinforcement learning towards self-improvement.

Two of the three reads approach the same capability from opposite ends: Learn2Play Bench measures how well LLM agents learn from experience in unfamiliar environments, while the MiMo-V2.6 technical report scales reinforcement learning toward self-improvement.

⚙️ Ship — 3 items

📦 Use Your Claude Plan in Amp — Claude Pro/Max subscriptions can now be used in Amp — free and available for everyone. Pick Mode > Claude Code in a new thread (it sits on the main dial, no trip into 'Other Agents') and link your Claude sub.

📦 OpenClaw 2026.9.9 — 185 commits · 112 pull requests · 92 contributors. Release notes and changelog carry the same content in two formats — expandable sections for people, and a plain-Markdown changelog formatted for AI agents and tools.

📦 Kiro — Workflow Feedback, Power Search, and Flexible Hook Matching — Kiro CLI 2.29.0 introduces Memory for local V3 sessions — Kiro learns useful context from your work and recalls it in later chat sessions, with controls for memory access and automatic updates.

💬 Community — 1 item

💬 Talorys – A self-hosted personal AI agent on Cloudflare's free tier — A self-hosted personal AI agent that runs on Cloudflare's free tier, code on github.com/rociiu — 161 points and 86 comments on HN, with commenters arguing over whether Cloudflare-hosted counts as self-hosted at all.

📚 Read — 2 items

📚 AgentGarten: Code Worlds for Evolving Agents — Agents learn through interaction, and what they can learn is bounded by the environments they learn in — AgentGarten's answer is code worlds built for evolving agents.

📚 DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks — Bootstraps agent memory from tasks the agent generates itself — memory/retrieval for agents, with a project page linked from the paper.

Both papers make the agent's own activity the source of its capability: AgentGarten holds that agents learn through interaction, bounded by the environments they learn in, and DAEDALUS bootstraps agent memory from tasks the agent generates itself.

💬 Community — 3 items

💬 Mxc: Microsoft Execution Containers version 1.0.0 — Microsoft Execution Containers reaches version 1.0.0 — 151 points and 33 comments on HN, where the top comment calls it 'finally answering bubblewrap' and takes up the industry's permissioning record.

💬 OpenAI annualised revenues $20B less than previously signalled — CNBC report that OpenAI's annualised revenues are $20B less than previously signalled — 429 points and 300 comments on HN, the day's biggest thread.

💬 Iranian campaign planted fake articles in real U.S. publications using ChatGPT — Washington Post report that an Iranian campaign planted fake articles in real U.S. publications using ChatGPT — 191 points and 172 comments on HN, with OpenAI's Thursday response quoted in the thread.

Both OpenAI threads arrive via third-party outlets — CNBC on the $20B annualised revenue gap, the Washington Post on ChatGPT-planted fake articles — and drew the day's two heaviest comment counts, 300 and 172.

📚 Read — 3 items

📚 SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles — Co-evolves agent skills and the agents that use them through dynamic skill lifecycles, with memory-augmented RL strengthening LLM agents' abilities.

📚 On KL-Regularized Policy Optimization — Paper on KL-regularized policy optimization (KLPO).

📚 VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction — Uses tool-augmented reinforcement learning to bridge causal transitions for video event prediction.

⚙️ Ship — 3 items

📦 MCP servers v1.0.0 — server-everything, server-filesystem, server-memory, and server-sequential-thinking publish at 1.0.0 on npm; mcp-server-fetch, mcp-server-git, and mcp-server-time publish at 2026.10.10 on PyPI. The everything server also cleans up subscriptions on session disconnect.

📦 OpenClaw plugin SDK moves SQLite runtime writes into workers — Plugin SDK refactor (#168857) moves SQLite runtime writes into workers — a change to where plugin runtime writes are executed; worth retesting any OpenClaw plugin setup that uses SQLite.

📦 New controls and chat improvements in Copilot for JetBrains — New controls and chat improvements in Copilot for JetBrains, with user experience enhancements, quality improvements, and a deprecation in the same changelog entry — check what's deprecated before updating.

💬 Community — 3 items

💬 Five months treating bugs like patients and coding agents like a medical team — Cockroach Labs writes up five months running coding agents as a medical team for bug triage — the piece says it consumed just over $135,000 in Claude tokens; 197 points / 94 comments on HN.

💬 OpenAI Withdraws 3 Math Papers — OpenAI withdraws three math papers — the HN thread is at 342 points but flagged as a duplicate, with comments moved to a separate discussion.

💬 My personal AI agent posted my bank details on company Slack — Business Insider first-person account: the author gave Grok Bot their financial details and connected it to work Slack, and the agent posted bank details to company Slack; 62 points / 66 comments on HN.

Two sides of giving agents real access: Cockroach Labs counts just over $135,000 in Claude tokens over five months of coding-agent bug triage, while the bank-details thread is a personal agent that had both financial details and work-Slack access posting bank details to company Slack.

📚 Read — 2 items

📚 Harness Engineering for Software Engineering via Modular Executable Dev-Primitives — HERMES turns repository components into agent-native Dev-Primitives — modular executable primitives, proposed as a harness-engineering approach for software engineering.

📚 Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight — Self-Retrospection Distillation turns post-hoc experiences into prior foresight — hindsight distilled into foresight available in advance. The author is answering questions on the paper page.

💬 Community — 4 items

💬 The gap in shared understanding of LLM capability is widening — HN thread (186 points, 285 comments) on Karpathy's argument that the gap in shared understanding of LLM capability is widening.

💬 OpenAI withdraws three mathematical results — HN thread (379 points, 609 comments) on OpenAI withdrawing three mathematical results, per the openai/math history.md; top comment asks whether the withdrawn proofs were the ones without Lean verification.

💬 500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter — HN thread (140 points, 103 comments) on using AI agents to decompile a first-person shooter after 500B tokens.

💬 @jwherrman.bsky.social: A lot of AI assistants seem to have been built around the assumption that the rest of the world is not actually real — John Herrman: "A lot of AI assistants seem to have been built around the assumption that the rest of the world is not actually real," linking to a New York Magazine piece on the AI agent flood.

The two highest-comment threads in the bundle both concern trusting LLM output: Karpathy's on the widening gap in shared understanding of LLM capability (285 comments) and OpenAI's withdrawal of three mathematical results (609 comments).

⚙️ Ship — 2 items

📦 Gemini 4 Argon spotted in A/B testing on Antigravity — Reportedly reaching some users through A/B tests on Antigravity; early outputs shared on X look like a clear step up from Gemini 3.8 Flash. Spotted in testing — not a confirmed launch.

📦 codex-profiles 1.3.0 — Adds a native profile and workspace menu with usage, a usage CLI command for remaining Codex quota, low-quota alerts, and a Raycast extension for profiles and workspaces. Maintainer waived Desktop smoke checks — real Desktop behavior is unverified.

📚 Read — 1 item

📚 TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models — Paper applies rollout-guided quantization-aware training to FP4 reinforcement learning of MoE language models.

⚙️ Ship — 2 items

📦 OMP 18.10.0 — Raises the default context window for openai-codex/gpt-6.1-sol and gpt-6.1-sol-wm from 272K to 500K, matching OpenAI's new Codex default; extended context still opens 922K. Also adds omp session pack/unpack/moved to move a session with its artifacts to another machine (resuming the copy left behind forks it), omp auth export/import to carry model sign-ins, and /fork to create a new Tern pane while a turn is streaming.

📦 Gemini 4 Argon hints emerge as Google tests Carbon checkpoint — Reportedly spotted in Antigravity: Gemini 4 Argon references with 256K, 512K and 900K context options, while Google employees reportedly test a separate Carbon checkpoint — no confirmed launch yet.

Both items put context-window scale at the center: OMP's new 500K default matches OpenAI's Codex default (extended context still opens 922K), and the reported Gemini 4 Argon options in Antigravity run 256K, 512K and 900K.

💬 Community — 1 item

💬 OpenAI, the Partition Principle, and Mathematics — HN discussion of OpenAI's math releases and the Partition Principle — 211 points, 313 comments. One framing in the thread: it can be true both that OpenAI is 'inconsiderate/harmful' with its math releases and that they have created 'a treasure trove'.

📚 Read — 2 items

📚 SWE-Duel: Towards a Self-Scaling Arena for Coding Agents — arXiv paper (submitted 8 Oct 2026) proposing SWE-Duel, a self-scaling arena for evaluating coding agents; authors Wei Minn, Zhensu Sun, Lwin Khin Shar, David Lo.

📚 A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization — Introduces AgenticBBO-Bench, a unified benchmark for LLM agents on black-box optimization tasks.

💬 Community — 1 item

💬 Codex CLI installs a persistent, self-updating background daemon by default — 425 points / 222 comments on HN. From the thread: there aren't any documented settings to disable this, but there is a feature flag it currently sits behind.