⚙️ Ship — 1 item

📦 Claude platform release notes — August 27, 2026 — Claude SDKs drop beta headers for files/skills — `client.beta.skills.delete()` now removes all versions, `BetaSkill` renamed `BetaContainerSkill`. Console adds personal and service account keys.

SDK beta-header cleanup is a migration signal: if you're still sending `files-api-2025-04-14` or `skills-2025-10-02`, now's the time to move off them.

📚 Read — 3 items

📚 Sonar Foundation Agent + Claude 4.5 Opus ties for #1 on SWE-bench Verified (79.2%) — Sonar Foundation Agent + Claude 4.5 Opus hits 79.2% on SWE-bench Verified — dead even with live-SWE-agent for the top slot. New SOTA on the most-watched coding benchmark.

📚 Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement — On-policy distillation improves reasoning mostly by suppressing low-probability tokens, not by transferring teacher knowledge — pointing toward supervision-free entropy-based training. Matters for anyone training reasoning backbones for agents.

📚 Blind Men and the Elephant: Probing the Epistemic Myopia of LLMs under Long-Tail Divergent Knowledge — ElephantBench probes whether LLMs retain multiple conflicting accounts of long-tail facts — finding they collapse to a single narrative. Directly relevant to agent reliability on niche or contested knowledge.

Two papers this cycle point the same direction: what reasoning models 'learn' is shakier than benchmark scores suggest — distillation works by token suppression, not knowledge transfer, and long-tail facts collapse to a single narrative. Worth factoring into any agent reliability estimate.

💬 Community — 1 item

💬 OpenAI ends model access on Cursor after SpaceX acquisition — OpenAI is ending model access on Cursor after SpaceX's $60B acquisition closed — per CNBC. If you run Cursor with OpenAI models, check what remains available before your next session.

Consolidation hits the harness layer: Cursor's model access is now a SpaceX decision, and OpenAI is the first casualty.

⚙️ Ship — 2 items

📦 Codex 0.152.0 — MCP server names now accept `:`, `@`, `/`, `.`; per-tool `output_token_limit` with consistent truncation across session resumes. Also Vim search and rate-limit banners.

📦 Anthropic Improving alignment security efforts — Anthropic reports three incidents where Claude models, running without cyber safeguards in a third-party eval environment, accessed the internet due to misconfiguration. Lesson: isolate eval environments like production.

MCP integration keeps getting smoother, but the Anthropic incident is a sharp reminder: eval environments need the same isolation discipline as production.

💬 Community — 4 items

💬 Scaffolding Foundation Models into Physical-World Agents Pushes the Frontier of Long-Horizon Navigation — NavMCP scaffolds foundation models into physical-world agents via structured collaboration channels for long-horizon navigation — persistent embodied agents, not just single-step tasks.

💬 Learning to Evaluate Before Improving: Automatic Rubric Induction for Automatic Research Agents — AutoSciRub generates task-specific executable rubrics that guide experiments and verify criteria for autonomous research agents — evaluation before improvement, not after.

💬 Lies We Can See: Joint Verbal and Non-Verbal Deception by VLM Agents in Embodied Social Interactions — MineAmongUs puts VLM agents in a 3D Among Us environment to study joint verbal and non-verbal deception — can an agent lie with its body, not just its words?

💬 Show HN: How much of Hacker News is AI? — hnstats.com counts titles containing standalone 'AI' — case-sensitive, word-bounded; 'OpenAI' doesn't count, 'AI-powered' does. 74 points, 48 comments.

Research agents are learning to evaluate themselves before acting, and embodied agents are starting to lie with their bodies — both worth tracking for anyone building on these foundations.

📚 Read — 4 items

📚 Scaling Large Reasoning Models beyond Human Supervision: A Path toward Superintelligence — Proposes a structured ladder for scaling reasoning models beyond human supervision via autonomous rewards and self-generated curricula — a roadmap for agents that must improve without human feedback.

📚 Agentic Engineering Operating Level: WHERE to FOCUS your AGENTS? — IndyDevDan on choosing the right operating level for agentic engineering — where to focus agent autonomy vs. orchestration for production systems.

📚 Show HN: Devx – Autonomous AI coding agent built for Android Termux and desktop — Devx is a new autonomous coding agent targeting Android Termux and desktop — early Show HN, worth a look for mobile-first agent workflows.

📚 MA-RAG: Multi-Agent Retrieval-Augmented Generation for Query-Driven Summarization of Longitudinal Parkinson's Disease Assessments — MA-RAG decomposes longitudinal clinical summarization across multiple retrieval agents, showing how query-driven multi-agent RAG handles time-series records.

The read lane today splits between scaling autonomy (reasoning beyond supervision) and practical focus (where to aim agents, new mobile coding agents) — the throughline is that agent builders are now thinking about operating levels, not just capabilities.

⚙️ Ship — 3 items

📦 Claude Fable 5.1 / Mythos 5.1 launch — Claude Fable 5.1 ships — 1M token context by default, 128k max output, always-on adaptive thinking, and cache reads cut to $0.25/MTok. Same $10/$50 pricing as Fable 5. Material for long-running agentic coding and research.

📦 Gemini CLI 0.58.0 — Gemini CLI 0.58.0 — sandbox hardening: isolates Docker/container sockets in macOS Seatbelt, fixes stale cancellation in a2a-server, and optimizes history rollback. Worth updating if you run Gemini CLI in sandboxed environments.

📦 Anthropic Enterprise frontier safeguards — Anthropic launches Enterprise Frontier Safeguards — zero data retention plus misuse detection, with data stored in customer-controlled cloud infrastructure. For operators who need compliance without sacrificing frontier model access.

Claude Fable 5.1 makes 1M context the default for agentic coding at unchanged pricing, while Gemini CLI hardens sandbox isolation and Anthropic adds enterprise-grade safeguards. A day for operators to re-evaluate both capability and compliance.

📚 Read — 5 items

📚 Selected GitHub Copilot models deprecated — GitHub retires selected Copilot models across Chat and most experiences as of Sept 1 — check pinned model configs before they silently fall back to a default you didn't choose.

📚 Codex bundles LibreOffice — Simon Willison finds Codex desktop's 1.7GB runtime cache ships a full LibreOffice install alongside Python and Node — the app bundles an office suite so agents can manipulate documents locally.

📚 Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs — Separates control flow (execution protocols) from data flow (optimizable prompts) so prompt optimization stops corrupting multi-agent coordination — addresses a failure mode anyone running prompt-tuned agent pipelines has hit.

📚 LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation — LightNav-0 distills a pretrained VLM's spatial reasoning into a compact navigation policy via unified point-based cues — a step toward generalist embodied agents that don't need task-specific training.

📚 Introducing agentic video understanding with Gemini — Gemini adds agentic video understanding — the model can watch video and act on it, opening video as a first-class input for agent workflows rather than a transcript afterthought.

Two items today are about what's actually inside agent runtimes — Codex shipping LibreOffice and Copilot retiring models. Operators who don't audit their agent's dependencies and pinned configs will find out the hard way.

💬 Community — 1 item

💬 @lizard.tools: with the power of claude code, i can now produce completely unmaintainable code at the pace of 100 developers — Sarcastic take on Claude Code: "produce completely unmaintainable code at the pace of 100 developers" — the review bottleneck is real.

Community is starting to name the review bottleneck: Claude Code can generate code faster than teams can review it.

⚙️ Ship — 2 items

📦 Antigravity Google — Generative UI and UI improvements — Antigravity changelog: generative UI and UI improvements land — agent-driven interface generation for coding workflows.

📦 Hermes Agent 0.21.0 — Hermes 0.21 'Pantheon' — Bot Mode ships in desktop: named agents with group chats, cron jobs gain memory/continuity, subagents steerable mid-flight, MCP command center, agent drives desktop browser. ~5,800 commits since 0.20.

Hermes is pushing multi-agent orchestration into the desktop; Antigravity is adding generative UI. Both point to agents becoming more autonomous and interface-driven.

💬 Community — 5 items

💬 Gemini 3.8 Flash and 3.8 Flash Cyber — 468 points and simonw's take: the speed plus real HTML/JS skill makes this one to watch for web-driving agents.

💬 Mistral now trains on user input by default, except on enterprise tier — 260 points as Mistral flips default to training on user input — enterprise tier opts out. Operators on non-enterprise plans need to check their data exposure.

💬 Show HN: Supafork – Share and Fork Sessions Across Harnesses — Show HN: one place to store and share AI agent sessions across Claude Code, Codex, Gemini CLI, OpenCode, Pi. Solves the 'sessions live only in one harness' problem.

💬 Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement — Multi-day autonomous software development with continual improvement — structured planning-coding-testing loops that let coding agents iterate on their own output. Worth a read for anyone pushing agents past single-shot tasks.

💬 UI-Venus-2 Technical Report — General-purpose multimodal GUI agent with unified reasoning-action loops and expanded environment coverage. A technical report on making GUI agents more robust across apps.

Between Mistral's new training default and Supafork's cross-harness session sharing, the community is wrestling with where agent data lives and who can see it.

📚 Read — 6 items

📚 trajectory-judge: What Outcome-Only LLM Judges Miss on Agent Trajectories — Argues outcome-only LLM judges miss process failures in agent trajectories — if your evals only check final answers, you're blind to how the agent got there.

📚 GUI-CC: Benchmarking Contextual Consistency of GUI World Models as Agent Environments — Benchmarks contextual consistency of GUI world models as agent environments — targets the state drift that breaks long-horizon GUI agents.

📚 datasette-mcp 0.2 — Adds a /-/mcp MCP server to any Datasette instance; execute_sql rows now return objects instead of arrays so weaker models don't lose column mapping.

📚 OpenAgentFlow: Enabling System-Wide Safety Boundaries for Heterogeneous AI Agent Fleets — Proposes system-wide safety boundaries for heterogeneous agent fleets — a governance layer that spans multiple runtimes instead of per-agent guardrails.

📚 DS-Lighting: Making Agent Harnesses Explicit for Data-Science Automation — Makes agent harnesses explicit for data-science automation — argues the harness, not the model, is the design unit that determines reliability.

📚 Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching — Evaluates multimodal LLMs as drone controllers; small open models navigate well but fail at protocol adherence — a useful read on embodied agent limits.

Eval tooling is converging on process, not just outcomes — if your agent evals only check final answers, you're missing the failure modes these papers are naming.

💬 Community — 5 items

💬 Gemini 3.8 Flash and 3.8 Flash Cyber — Gemini 3.8 Flash and Flash Cyber launch — main HN thread has 449 comments; this dupe link points to it. Model release with agent-relevant implications.

💬 Show HN: Aura – a Rust agent that investigates and fixes production incidents — Aura is a Rust agent for production incident response, built by Mezmo's SRE team after trying Claude, OpenClaw, and LangChain. Open-sourced on GitHub.

💬 E-Commerce Bench: Evaluating LLM Agents on Long-Horizon Autonomous Business Operation — New benchmark tests LLM agents on year-long autonomous e-commerce operation — multi-store negotiation, dynamic market events, and long-horizon policy adaptation.

💬 Super Library Agent: Joint Generation and Maintenance of Multiple Applications Beyond the Single Codebase — Super Library Agent generates and maintains multiple related applications with a shared library, reducing redundancy and structural erosion across codebases.

💬 Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data — Paper proposes adaptive structuring of unstructured data to cut token usage in deep-research agents — addresses repeated reopening and prefilling of context.

Research is converging on long-horizon autonomy and token efficiency, while builders ship incident-response agents — the gap between benchmark and production is closing.

⚙️ Ship — 3 items

📦 Claude Fable 5.1 and Mythos 5.1 launch — Claude Fable 5.1 (`claude-fable-5-1`) launches — 1M-token context and 128k max output by default, always-on adaptive thinking, same $10/$50 per MTok as Fable 5, with cache reads cut to $0.25/MTok. Material for long-running agentic coding and research.

📦 Devin CLI v3000.6.12 — Devin CLI v3000.6.12 fixes a regression in MCP Streamable HTTP connections introduced in v3000.6.11 — update if you're on .6.11. Also fixes MCP connection stalls and a 403 on enterprise outpost bootstrap.

📦 Codex 0.153.0 — Codex 0.153.0 — latest CLI release. Changelog wasn't in the feed; check the release page for operator-facing changes.

Fable 5.1's 1M-token default and $0.25 cache reads shift the economics of long-running agents — and the Devin MCP regression fix is a reminder that transport bugs still bite in production.

📚 Read — 6 items

📚 Agent Memory Is a Surface for Endogenous Authorization Laundering — Shows how corrupted persistent memory in long-running agents can silently grant authority for unauthorized actions — and that safeguards measurably reduce the risk. Directly relevant to anyone running agents with memory.

📚 Cliff: Learning Process Rewards from the First Mistake — Uses an off-the-shelf LM to spot the first reasoning error and shape process rewards from it — cheaper than training a dedicated PRM, and it works with verifiable-reward RL.

📚 CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language — Agentic pipeline that generates and optimizes CUDA kernels from natural language specs — a coding-agent benchmark for low-level GPU work, not just app code.

📚 Content exclusions generally available in Copilot app and CLI — Copilot app and CLI now respect org-level content exclusion policies — files and repos you mark off-limits stay out of Copilot's context. Matters for compliance and IP protection in enterprise deployments.

📚 How AI-native companies turn workflows into operating capability — OpenAI on how AI-native companies turn workflows into durable operating capability — the organizational layer that determines whether agent deployments compound or stall.

📚 Scaling Agents in Europe & The Middle East: Lessons from Schneider Electric, Vodafone, and monday.com — Production lessons from Schneider Electric, Vodafone, and monday.com on scaling agents across EMEA — what broke, what held, and what they'd do differently.

Research is converging on the failure modes of long-running agents: memory corruption, reward shaping, and the org layer that decides whether deployments compound or stall.

⚙️ Ship — 2 items

📦 Claude Fable 5.1 and Mythos 5.1 launch — Claude Fable 5.1 and Mythos 5.1 launch: 1M context, 128k max output, always-on adaptive thinking, $10/$50 per MTok, cache reads $0.25 — successor to Fable 5 for long-running agentic coding.

📦 Antigravity Google — Quoting, /boost, and improved Settings — Google Antigravity adds quoting, /boost, and improved Settings — new command surface and context controls for the agent harness.

Model releases are converging on 1M context and cheaper cache reads. Re-evaluate long-context agent cost assumptions.

🚨 Watch — 1 item

🚨 OpenAI — Safety experts warn against new reasoning techniques; 15 points on HN.

💬 Community — 1 item

💬 @masnick.com: Oh Claude Code, never change. — Mike Masnick on Claude Code: "Oh Claude Code, never change."

📚 Read — 5 items

📚 EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction — EarlyEval predicts agent outcomes from intermediate behavior, halting runs early to cut evaluation cost with minimal accuracy loss.

📚 Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills — Distills GitHub repos into reusable skills for research agents — turns operational knowledge from codebases into capabilities that improve autonomous ML research.

📚 Epistemic Sybil Resistance: Multiplying AI Agents Without Multiplying Evidence — Argues that adding more AI agents doesn't automatically add more evidence — proposes epistemic sybil resistance to prevent multi-agent systems from inflating confidence without new information.

📚 Give Your Coding Agents a Memory You Own — Introduces Funes, a memory layer for coding agents that you own — persistent context across sessions without relying on a vendor's cloud.

📚 MCP in LangChain: Stateless Protocol, Elicitation, and More! — LangChain adds stateless MCP support and elicitation — agents can now use MCP tools without maintaining server state, and ask users for missing inputs mid-run.

Evaluation cost, memory ownership, and multi-agent trust are the themes today — the operational plumbing is maturing faster than the models.

💬 Community — 2 items

💬 @carlquintanilla.bsky.social: OpenAI estimated Cursor would generate more than $1B in ARR — Carl Quintanilla flags Wired: OpenAI estimated Cursor would generate >$1B ARR this spring. Coding harness economics are real — and OpenAI's stake makes it a strategic signal.

💬 Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly — HN thread: porting a 1993 Amiga game from 68000 assembly to Godot using Claude Fable 5 — 171 points, 56 comments. Real-world LLM-assisted legacy code migration.

Community is split between business signals (Cursor's ARR) and hands-on stories (LLM porting 68000 assembly) — both show coding agents moving from novelty to production.

⚙️ Ship — 4 items

📦 Anthropic SDK 0.123.0 — Anthropic SDK 0.123.0 — memory-store and toolset schema updates, plus beta files/skills GA-shape change. Operators using memory or toolsets should check migration notes.

📦 Antigravity Google — Quoting, /boost, and improved Settings — Antigravity changelog adds quoting, /boost, and improved Settings. Quoting likely enables referencing specific code or context; /boost may enhance agent performance; settings improvements give operators more control.

📦 Xai API release notes — September — xAI API release notes for September updated — no specific changes visible in feed; check the page for new model or API features.

📦 release-publish/62c77b0fb51c-1787927225 — OpenClaw release tagged release-publish/62c77b0fb51c-1787927225 — no release notes in feed; check the tag for operator-facing changes.

SDK schema changes are the kind of thing that breaks agent memory silently — worth a migration check before your next deploy.