📚 Read — 1 item

📚 OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software — Extends OSWorld to scientific software — tests whether computer-use agents can learn and operate domain tools, not just navigate generic UIs. Scientific software presents a demanding test.

Benchmarks are moving from generic UI navigation to domain-specific software — expect agent evals to get harder and more specialized.

💬 Community — 1 item

💬 Show HN: Pi pod – Run your pi coding agent in sandboxes on your own server — Pi pod runs the Pi coding agent in isolated sandboxes ("pods") on your own server, with composable environments — a self-hosted path for teams moving agent workloads off SaaS.

Self-hosted agent sandboxing is becoming productized; if you're not ready to hand agent workloads to a SaaS, this is the pattern to watch.

⚙️ Ship — 2 items

📦 Copilot code review: API support and new default effort level — Copilot code review is now callable via API, and Balanced becomes the default effort level — teams can trigger reviews in CI without the UI.

📦 Factory 0.233.0 — Factory 0.233.0 — resume a session with a message, script output trimming keeps start/end, cut-off replies retry. Fixes oversized requests and malformed tool args.

Today's harness moves are about session continuity and context-window handling: resume-with-message and oversized-request fixes in Factory, API-driven review in Copilot. Operators should test these paths before relying on them.

📚 Read — 2 items

📚 BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation — Online Bayesian reliability memory that learns which agents to trust for consultation, adapting as reliability shifts. Useful for multi-agent routing where trust calibration matters.

📚 Show HN: Corral – Kill every command your agent starts — Kills orphaned processes agents leave behind — tail -f, background jobs — after the agent exits. Solves a real operational headache for anyone running agents that spawn subprocesses.

Process hygiene and trust calibration are the unglamorous edges of agent operations — today's items address both.

⚙️ Ship — 4 items

📦 release-publish/004549970362-1790888386 — New release
📦 Pinned inputs 9 — New Hermes release
📦 Copilot Dynamic workflows in CLI and the app — Copilot changelog update
📦 Kiro — Workflow Delegation, Live Steering Context, and Saved Prompt Commands — Changelog page updated

💬 Community — 2 items

💬 Inside Anthropic's quest to instill morality into its models — 357-comment HN debate — NYT details Anthropic bringing religious scholars in to help encode values into Claude; the 144-point HN thread skews hard skeptical ('bubble-phase research'), but this values-tuning is the layer that decides how Claude agents behave on ambiguous requests.

💬 PPSSPP forks multiplying 'given the turn toward Claude Code' — Retro-emulation dev flags the new PPSSPP forks (tico-ppsspp, ppsspp-wasm) and predicts more 'given the turn toward Claude Code' — agent harnesses making a complex C++ emulator maintainable by small teams. Thin engagement, but it's the agents-reshape-OSS-maintenance thesis in the wild.

The values Anthropic bakes into Claude decide how your agents handle ambiguous requests, and Claude Code is quietly making 'fork it and maintain it yourself' viable for big codebases — two capability shifts hiding in culture-war clothing.

📚 Read — 3 items

📚 EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments? — EngiWorld bills itself as the first benchmark covering the complete professional engineering workflow — not isolated coding tasks. Useful for calibrating what frontier agents actually deliver on multi-step engineering work.

📚 OMP 18.6.1 — Subagent follow-up messages now arrive as background jobs you can wait on or cancel (delivered once); fixes OpenAI context compaction wrongly rejecting sessions with many screenshots, and preserves DeepSeek-family reasoning context for warm Responses sessions.

📚 Ordewell 0.6.3 — Closes the planner confinement hole from 0.6.2 — a `cd` chain no longer lets the research agent read outside the workspace unprompted. Terminal transport now runs OpenCode 2.x and works on macOS without tmux.

Two runtime fixes are the actionable part today — Ordewell closed a workspace-confinement escape and OMP made subagent replies waitable background jobs — while the benchmark paper suggests frontier agents still don't clear end-to-end engineering work.

💬 Community — 2 items

💬 Agents don't need memory, they need documentation — Argues agents need documentation, not memory systems — 343 points / 210 comments on HN, with the top comment calling third-party memory systems 'LLM rube goldberg machines' that 'just pollutes context.'

💬 Greg Kroah-Hartman – Security in the LLM Age [video] — 'Security in the LLM Age' from Kernel Recipes 2026 — 336 points / 128 comments on HN, where the top comment transcribes slides from the talk.

The top comment rejects the essay's remedy too: 'You don't need documentation or the 3rd party memory systems. The code IS the documentation.'

⚙️ Ship — 2 items

📦 Pi 1.0.0 — The TUI runs fullscreen by default (`tuiMode: "regular"` keeps the terminal's normal scrollback); codemode uses about 40% fewer prompt tokens and its errors tell the model how to recover; MCP OAuth hardening adds `oauth.authServerMetadataUrl`, RFC 9207 `iss` checks, credentials per server, and step-up sign-in that keeps granted scopes. Also in 1.0.0: `models.generateImages()` in codemode using the session's credentials, and Radius sign-in plus its MCP server setup in one step from `/login`.

📦 Anthropic: Claude Frontier Academy — $100M to train 10,000 engineers — Anthropic commits $100M to train 10,000 Frontier Deployed Engineers and tackle the enterprise AI talent gap — trained to "the same standard of skills as Anthropic's own engineers," with first cohorts from Accenture, Bain, and Capgemini.

Pi's release notes pair the codemode token cut — about 40% fewer — with errors that tell the model how to recover.

📚 Read — 5 items

📚 Runmaestro/Maestro 0.17.8 — Security release: the built-in server that maestro-cli and the phone interface drive listened on the whole network from the moment Maestro opened, even with Live off, and answered requests from any web page — anyone with its URL could run commands. It now listens on 127.0.0.1 only until Live is turned on.

📚 fx 0.0.13 — Ultrafast lands on supported OpenAI models through AI Gateway (a higher-priced inference tier) — launches up to 23× faster, shell calls up to 8.6× faster, exits up to 44× faster in matched benchmarks; subagents and ACP sessions support it. Breaking: terminals launched through the shell tool now stop when fx exits, and new libfx checkpoints cannot be restored by older versions.

📚 Show HN: Offrun – manage every coding agent from one workspace — Run Claude Code, Codex, AGY, and Grok Build side by side — see who is working, who needs you, and what every account has left. 76 points and 64 comments on HN.

📚 1Jehuang/Jcode 0.90.1 — OpenAI-compatible providers that answer 529 or another 5xx "heavy usage" error are now retried automatically instead of failing the turn, and /usage shows real quota windows for Kimi Code, Cursor, and the Z.ai Coding Plan.

📚 EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making — HF Daily Papers selection on eliciting world knowledge in agents for transferable decision-making.

The Maestro release notes say this is a security release and every user should install it.

⚙️ Ship — 2 items

📦 Devin CLI v3000.11.1 / v3000.11.3 — v3000.11.1 adds `devin --cloud` / `/cloud` to run Devin Cloud sessions from the CLI, `/handoff` and `/pickup` to bring a cloud session's PR branch into a local session, and `devin ssh` / `/ssh` to connect to a Devin Cloud VM; ACP clients can now select Devin models, thinking effort, and speed. Agent events and metrics export to an OpenTelemetry collector via the `otel` config block or `OTEL_EXPORTER_OTLP_*` variables. v3000.11.3 fixes browser login to record the signed-in tenant's webapp host so `/handoff` and `/share` links open on `<customer>.devinenterprise.com` instead of `app.devin.ai` — re-run `/login` to pick it up.

📦 xAI API release notes — October — xAI API release notes page updated for October; the docs list Grok 4.7 under Latest Models and include a Multi Agent section.

The Devin CLI changelog lists v3000.11.1 (September 21, 2026) and v3000.11.3 (September 22, 2026); the v3000.11.3 browser-login fix requires re-running `/login` to take effect.

💬 Community — 1 item

💬 Getting the most out of Opus 5.5 in Claude and Claude Code — claude.dev post on getting the most out of Opus 5.5 in Claude and Claude Code — 238 points and 157 comments on HN; a commenter describes giving Opus general directives to speed up CI while caring about billing minutes.

One commenter reports giving Opus general directives to speed up their CI and telling it they care about billing minutes.

📚 Read — 5 items

📚 Terraform for agents: how we define our software factory in code — Warp's post on defining its software factory in code, Terraform-style — the pattern behind Warp Factories, which run fleets of coding agents across the SDLC (first-pass review on every PR, bug reproduction and routing).

📚 Fast Models, Slow Evidence: A Paired and Self-Audited Evaluation of System-1 Decision Models for LLM Agent Harnesses — Paired, self-audited evaluation of System-1 decision models for LLM agent harnesses — the 'fast models, slow evidence' question; arXiv 2610.02267, submitted 1 Oct 2026, single author (Jiawei Li).

📚 Choosing Before Acting: Comparative Value Estimation for Long-Horizon Tool-Use Agents — Comparative value estimation for long-horizon tool-use agents — choosing before acting; arXiv 2610.02330, submitted 1 Oct 2026 (Yu Li, Zheng Zhang, Xin Liu, et al.).

📚 DAGent: Evaluate-then-Grow Planning for Deep Research Agents — Evaluate-then-grow planning for deep research agents; the authors' note on the paper page marks it as a NeurIPS 2026 paper.

📚 Sharpening Tax in Post-Training — Top-scored paper in today's HF Daily Papers batch — post-training work on the 'sharpening tax'; abstract is on the paper page.

Both arXiv papers were submitted on 1 Oct 2026 and both concern agent decision-making — fast System-1 decision models in harnesses, and comparative value estimation before tool-use agents act.

💬 Community — 3 items

💬 Anthropic reported diary entry to police, woman faces felony charge — Anthropic reported a diary entry to police and the woman faces a felony charge — 492 points / 422 comments on HN; discussion in the thread cites Florida Statute 836.10, which makes it a second-degree felony to send, post, or transmit a written or electronic record threatening...

💬 OpenAI "rogue" agent activities found on Wikimedia projects — OpenAI "rogue" agent activities found on Wikimedia projects — the HN submission links to wikimedia.org as the source; 254 points / 171 comments.

💬 Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates — Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates, per vals.ai — 178 points / 136 comments on HN.

The two most-engaged threads both involve agent providers and outside parties — Anthropic reported a diary entry to police, and OpenAI "rogue" agent activities were found on Wikimedia projects.

⚙️ Ship — 3 items

📦 Claude platform release notes — October 1, 2026 — Models API adds a `line` field: `GET /v1/models` and `GET /v1/models/{model_id}` now return the model line (Opus 4.5 and 4.6 both report `opus`; null for a model in no line). Same page: Claude Sonnet 4.5 (`claude-sonnet-4-5-20250929`) is deprecated, retirement on the Claude API scheduled for November 30, 2026 — migrate to Claude Sonnet 5.5.

📦 Devin: October 2, 2026 — Oct 2 notes: sidebar search now surfaces matching sub-sessions and workflow sessions under collapsed parents, and the session status stops prompting to review/approve/merge a PR once that step is done. Long-session fixes cover transcript flicker/blanking on background-tab return and self-scrolling while reading older history.

📦 Pinned inputs 8 — Hermes Agent release 'Pinned inputs 8' — no release notes in the feed; the tag is the only source.

The same Claude platform page adds a `line` field to `GET /v1/models` and `GET /v1/models/{model_id}`; Opus 4.5 and 4.6 both report `opus`, and `line` is null for a model that belongs to no line.

📚 Read — 4 items

📚 OmniSeek: Native Tool Integration for Multi-turn Audio-Visual Reasoning — Adobe paper on native tool integration for multi-turn audio-visual reasoning; the page notes code, weights, and data are under Adobe's internal review.

📚 4 new model(s) priced in the LiteLLM registry — Four new models added to LiteLLM's pricing and context registry: chatgpt/gpt-6-astra, chatgpt/gpt-6-luna, chatgpt/gpt-6-sol, and chatgpt/gpt-6.1-sol.

📚 Socai-Io/Socai 0.6.4 — Adds per-platform browser pacing, cloud session workers, and click-first deep reads on X; telemetry errors now redact provider API key values.

📚 The-Open-Engine/Zeroshot 10.10.1 — Preserves the invoking shell's USER in local Claude processes so macOS Keychain login works without a wrapper; hosted processes still ignore native context.

⚙️ Ship — 2 items

📦 Kiro — Workflows, Safer Untrusted Workspaces, and Enterprise Sign-In Controls — Claude Sonnet 5.5 is now available with experimental support across the Kiro IDE, CLI, Crew, and Web for Pro, Pro+, Pro Max, and Power customers in us-east-1 (IAD) and eu-central-1 (FRA); the Oct 2, 2026 changelog update also ships Workflows, safer untrusted workspaces, and enterprise sign-in controls.

📦 Devin: October 5, 2026 — Session workspace gets a Terminal tab for running commands on Devin's machine — the Files tab can now be opened while Devin is asleep, and opening Files or Terminal wakes Devin. Also in this cut: a ⌘K command palette, Devin Review effort levels, and Slack user-group mentions rendering as named @group mentions in the webapp.

Kiro's Sonnet 5.5 support is labeled experimental; Devin's new Terminal tab runs commands on Devin's machine, and opening Files or Terminal wakes a sleeping session.

💬 Community — 3 items

💬 Mistral Large 4 — Mistral Large 4 discussion — 1037 points / 680 comments on HN; model card at docs.mistral.ai/models/mistral-large-4-0.

💬 An AI agent emailed researchers for help. It told us why — science.org report on ColonistOne, an agent given the task of 'telling more humans about its work' — 50 points / 80 comments on HN (flagged).

💬 Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation — HF Daily Papers: causal decomposition finds self-generated feedback destabilizes test-time training in long-horizon adaptation.

Mistral Large 4 drew 680 comments on HN; the ColonistOne agent story drew 80.

📚 Read — 5 items

📚 ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience — ASCENT proposes online test-time training for long-horizon agents via self-distillation of verified experience; the page excerpt says it lets an LLM agent keep learning.

📚 SearchJev: A Fast and Calibrated System-1 Model for Search Agents — SearchJev is presented as a fast, calibrated System-1 model for search agents.

📚 Beyond Semantic Similarity: Performance and Costs of Agentic Retrieval for Complex Tasks — Compares the performance and costs of agentic retrieval against semantic similarity for complex tasks.

📚 Advancing computer use with Ironclad — OpenAI post on advancing computer use with Ironclad.

📚 Introducing Mistral Large 4 — Mistral announces Mistral Large 4.

ASCENT's excerpt says it lets an LLM agent keep learning; SearchJev is described as a fast, calibrated System-1 model for search agents.

🚨 Watch — 1 item

🚨 Code scanning AI Scan enablement status in security overview — Organization and enterprise administrators can now see AI Scan for pull requests enablement status in security overview. Filed under Improvement in the October 6, 2026 changelog.

The changelog entry is filed under Improvement (October 6, 2026): organization and enterprise administrators can now see AI Scan for pull requests enablement status in security overview.

💬 Community — 2 items

💬 Sharing AI Progress in Mathematics — OpenAI's post on AI progress in mathematics, 244 points / 190 comments on HN; links an openai/math repo whose preprints include a proof of Barnette's Conjecture.

💬 OpenTPU – An open-source AI accelerator, developed by AI — HN front page at 210 points / 275 comments on an open-source AI accelerator from github.com/fesens; top comment asks why the labs aren't burning their frontier models into chips already.

The OpenTPU thread's top comment asks why frontier labs aren't burning their frontier models into chips already.