📚 Read — 4 items

📚 How Third Moment Research runs 10,000 agents for 36+ hours — OpenAI's account of Third Moment Research running 10,000 agents for 36+ hours.

📚 Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context — LangChain on context engineering with Deep Agents — how to remember without keeping everything in context.

📚 The Most Common Questions About Agent Evals, Answered — LangChain's answers to the most common questions about agent evals.

📚 Amp puts AI agents to work beyond code with OpenAI models — OpenAI on Amp putting AI agents to work beyond code with OpenAI models.

💬 Community — 1 item

💬 OpenAI engineer arrested in bugging of executive conference room — SF Standard reports an OpenAI engineer was arrested for bugging an executive conference room — 646 points and 532 comments on HN, with a post from the OpenAI Newsroom account linked in the thread.

⚙️ Ship — 3 items

📦 Devin CLI v3000.11.92: /goal command and mid-turn steering — New /goal <objective> command keeps Devin working across turns until it reports the goal complete or blocked — a run pauses itself after 50 turns, not available on prompt-based plans. Steering hands mid-run messages to the running turn instead of the queue; new config.json settings include agent.subagent_model and local_tools_bypass_permissions (starts local-tools connections with Bypass Permissions on).

📦 Copilot: Agentic autofix for code scanning alerts general availability — Agentic autofix for code scanning alerts is now generally available — the changelog entry covers how it works, how to use it, access, and billing.

📦 Devin: October 14, 2026 — Interrupting Devin keeps the partial reply in chat, browser previews in Slack-started sessions post an "Open preview" link to the thread, and admins can assign a new org-level Review Manager role.

Both Devin updates change mid-turn control the same day: the CLI hands a message typed while Devin is working to the running turn instead of the queue, and the web product keeps the partial reply you already saw when you interrupt.

🚨 Watch — 1 item

🚨 Claude Code Desktop Vulnerability: Preview Pane Data Exfiltration — noma.security reports a data exfiltration vulnerability in Claude Code Desktop's preview pane; the HN thread — 15 points, no comments — hasn't surfaced any fix or CVE details yet.

💬 Community — 1 item

💬 OpenCode 2.0 — 812 points / 350 comments on HN — discussion covers usability, the client/server model, and OpenCode's ability to edit its own config and write plugins.

📚 Read — 4 items

📚 How Netflix Made AI Coding Agents More Efficient with JetBrains IDE Intelligence — JetBrains case study on how Netflix used IDE intelligence to make AI coding agents more efficient — relevant if your coding agents run inside a JetBrains IDE.

📚 Accelerating the search for antimicrobial peptides with Google Antigravity — DeepMind post on using Google Antigravity to accelerate the search for antimicrobial peptides — the coding harness applied to a scientific research workload.

📚 RuleTrek: Benchmarking Agentic Rule Reasoning in Complex Scenarios — New benchmark paper measuring agentic rule reasoning in complex scenarios — a fresh eval target for agent reasoning.

📚 ActionOPD: Distilling Decisions, Not Tokens — Introduces ActionOPD — a distillation approach that targets decisions rather than tokens.

Both fresh vendor posts are harness-in-action accounts: Netflix using JetBrains IDE intelligence to make coding agents more efficient, and DeepMind using Google Antigravity to accelerate antimicrobial peptide search.

💬 Community — 1 item

💬 A look at the math that OpenAI claims computes faster Fourier transforms — A close look at the math behind OpenAI's claimed faster Fourier transforms is drawing heavy HN discussion — 245 points and 168 comments on the thread.

⚙️ Ship — 3 items

📦 Codex 0.163.0 — parallel read-only tools, mid-response steering — Independent read-only tools (skill, memory, conversation-history lookups) now run in parallel, and a follow-up sent while Codex is responding steers it immediately — interrupting a turn also stops active Code Mode cells and their nested tool calls. Also: /config preference search, managed background-server restarts from /daemon, a /keymap leader key (default Ctrl+X, voice mute moves to F9), and Daybreak enabled by default for eligible accounts.

📦 Kiro: Kimi K3 now available — Moonshot AI's open-weight Kimi K3 lands with experimental support across the Kiro IDE, CLI, Crew, and Web, for Pro, Pro+, Pro Max, and Power customers on Kiro's US geography. Accepts text and images, with a 1M-token context.

📦 Claude: Code review is now generally available — Anthropic moves code review to general availability, per the announcement on Anthropic's own site.

Codex 0.163.0 turns Daybreak on by default for eligible accounts the same day Kiro adds Kimi K3 with experimental support — both of today's harness releases widen model access.

🚨 Watch — 2 items

🚨 Claude Code: plugin hash pin bypass (CVE-2026-86063) — CVE-2026-86063 — a plugin hash pin bypass in Claude Code: the hash pinning applied to plugins can be circumvented. Rated high severity; check the advisory if you rely on hash-pinned plugins.

🚨 Cursor: Elevated errors affecting Grok models — Some Grok-model requests in Cursor fail with a capacity error; a fix was implemented Oct 15, 2026 - 22:58 UTC and the incident is in monitoring. Retrying the request or switching to another model may help.

📚 Read — 4 items

📚 Athanor: An Agent-Native System for Autonomous Post-Training Research — Agent-native system in which Claude Opus 5.5 agents post-trained Qwen3-4B-Base on their own — autonomous post-training research without a human in the loop.

📚 EdgeWise: Bringing Long-Horizon Agents to the Edge via Selective Cloud Delegation — EdgeWise pairs a local small language model with selective cloud delegation via reinforcement learning to cut inference cost on long-horizon agent tasks.

📚 The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery — Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn et al. orchestrate coding agents for automated scientific (co-)discovery; submitted to arXiv 13 Oct 2026.

📚 Linus Torvalds Working On A Hefty Linux Kernel Improvement With Claude Code — Phoronix reports Linus Torvalds is working on a hefty Linux kernel improvement using Claude Code.

Athanor and The Agentic Scientist both put agents in the researcher's seat — one autonomously post-training Qwen3-4B-Base, the other orchestrating coding agents for scientific (co-)discovery.

⚙️ Ship — 2 items

📦 Factory CLI v0.243.0 · Desktop v0.200.0 — Droid now reads a range of a very large file without running out of memory, keystrokes reach the browser pane they target, and a background command that deletes a file with rm -f is no longer refused. A review workflow required across an organization now runs Droid on new pull requests, and an attachment still uploading can be canceled from the chat box.

📦 Multi-repo support and other improvements for GitHub Copilot in Slack and Microsoft Teams — Multi-repo support for GitHub Copilot in Slack and Teams, plus other improvements for Copilot in those surfaces — changelog entry dated October 15, 2026.

Both entries widen agent scope past a single repository — Copilot in Slack and Teams adds multi-repo support, and Factory's organization-required review workflows now run Droid on new pull requests.

💬 Community — 2 items

💬 More on Anthropic's Scion ripping off Mononoke — HN thread (249pts / 82 comments) on Anthropic's Scion and Meta's Mononoke; the original blog post author replies in-thread, saying the credit to Sapling follows how Meta itself externally refers to it.

💬 Show HN: Gemini 4 Argon plays StarCraft — Show HN: StarSkirmish Bench, a benchmark where LLMs play StarCraft: Brood War by writing code, refreshed with results from the flagship models of OpenAI, Anthropic, Google, and Muse (107pts / 51 comments).

📚 Read — 3 items

📚 Data World: Scaling Data Environments from Scratch — A multi-agent engine synthesizes data environments and tasks from scratch to train models for analytical insight discovery, targeting a common pain point in the data-agent space.

📚 MMSearch-OPD: Easy Made Hard, Hard Made Learnable for Long-Horizon Multimodal Search Agents — Refines existing easy VQA data into harder long-horizon tasks and trains multimodal search agents with reinforcement learning plus on-policy distillation.

📚 Player-1: Learning to Play from Synthetic Experience — Player-1 learns to play from synthetic experience.

Both Data World and MMSearch-OPD target agent training data: Data World synthesizes data environments and tasks from scratch, while MMSearch-OPD refines existing easy VQA into harder long-horizon tasks.

🚨 Watch — 1 item

🚨 Claude Agent SDK: Argument Injection via resume Option Allows Arbitrary Command Execution (CVE-2026-96620) — CVE-2026-96620 — argument injection through the resume option allows arbitrary command execution in claude-agent-sdk. Anyone using resume should check the advisory for affected versions and update.