📚 Read — 2 items

📚 Copilot code review: Resolution reasons and expanded capabilities — Copilot code review adds comment resolution reasons and expanded capabilities — reviewers can now record why a thread was resolved, making review outcomes auditable.

📚 CIFQA: A Deterministic Tool-Grounded Multi-Agent LLM Framework for Financial Query Answering — CIFQA: deterministic tool-grounded multi-agent framework for financial QA — claims deterministic outputs by grounding each agent in tools; relevant if nondeterminism is a liability in your multi-agent stack.

Deterministic tool-grounding is the throughline: CIFQA applies it to financial QA, and Copilot's resolution reasons make review outcomes auditable — both push toward less guesswork in agent outputs.

⚙️ Ship — 2 items

📦 Codex 0.151.0 — Codex 0.151.0 — extensions can now inspect or replace MCP tool results before they reach the model, nested subagent token usage counts toward root goal budgets, and `/cd` no longer weakens sandbox restrictions.

📦 OpenHands 1.16.0 — OpenHands 1.16.0 — skills move from an all-on catalog to an explicit allow-list, LLM provider selection lands in settings, and the Canvas Extensions frontend ships. Self-hosted backends get the VSCode button fix.

Sandbox and MCP control are today's theme: Codex hardens both, OpenHands moves skills to an explicit allow-list. Operators get more control over what agents can touch.

💬 Community — 1 item

💬 OpenAI pulls its models from Cursor over Musk contract concerns — OpenAI says its models will no longer be available in Cursor, citing "our experience with Elon Musk's companies violating contracts." A material shift for one of the most-used coding harnesses — Cursor operators need to plan around losing GPT access.

If you run Cursor in production, today is the day to audit which model your workflows actually depend on — GPT access is going away.

📚 Read — 2 items

📚 Build an Auditable VC Research Agent with Perplexity — LangChain shows how to build a VC research agent with Perplexity's Agent API, LangGraph, and LangSmith — every step traced for auditability.

📚 How to design an Agent for Production — LangChain's production agent design guide — covers observability, evaluation, and deployment patterns for agents that need to survive real traffic.

Two practical LangChain pieces on building agents that can be audited and run in production — the operational side of agent engineering is getting more concrete.

⚙️ Ship — 2 items

📦 Anthropic SDK 0.121.0 — Adds Standard Schema support for structured outputs and tools (type-safe tool definitions via Zod/Valibot) plus a beta `updates` thinking display mode.

📦 OpenAI SDK 3.5.0 — Makes function call output call IDs optional — less boilerplate when parsing tool results.

Both major SDKs shipped ergonomic improvements around tool calls today — worth updating if you're building agent loops.

📚 Read — 2 items

📚 Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models — Uses agentic game development to generate verifiable long-horizon trajectories for post-training spatial world models — game engines double as executable checkers for RL data.

📚 LLMs for Academic Workflows: An Evaluation of Literature Reviews Generated with Short and Long Context Windows of LLMs — Compares short vs long context windows for LLM-generated literature reviews — directly relevant to context-budget decisions in long-horizon agent workflows.

World-model training is borrowing agentic pipelines for data generation; expect more agent-generated, verifiable trajectories in post-training.

⚙️ Ship — 2 items

📦 Anthropic SDK 1.1.0 — Adds `updates` thinking display mode (beta) and Organization API endpoints; fixes tool runner to keep going on `pause_turn`. Worth updating if your agent loops pause mid-turn.

📦 OpenAI SDK 3.6.0 — Adds `compute_units` to Responses and Chat Completions usage for compute-based billing; hardens X.509 workload identity auth. Mostly deps otherwise.

Two SDK updates today: Anthropic's pause_turn fix matters for agent loops, OpenAI's compute_units gives you a new billing metric. Small but worth pulling.

💬 Community — 4 items

💬 WeMM-Embedding — Technical report on multimodal embedding models aligning text, images, and video in a shared space.

💬 Video-IFBench — A benchmark for measuring instruction following in multimodal LLMs across video scenarios.

💬 Code World Model — Paper proposing a coding agent architecture that separates world dynamics from visual rendering.

💬 hailey.at — Vague signal on Bluesky (54♥ / 5🔁) suggesting imminent movement.

📚 Read — 3 items

📚 FrontierChallenge: Evaluating Scientific Workflow Completion — FrontierChallenge evaluates end-to-end scientific workflows across domains; frontier models complete only ~20% of tasks — a hard number on how far agents are from autonomous science.

📚 Standalone LLM and a Pre-specified Agentic Pipeline for Explaining ICU Mortality Predictions: a Feasibility Study on the eICU Demo Dataset — Feasibility study on eICU: a pre-specified agentic pipeline beats a standalone LLM at explaining ICU mortality predictions — early evidence that structured agent workflows outperform raw prompting for clinical explainability.

📚 A smarter way to run code migrations with less LLM context — Sourcegraph's Deep Search runs migration scripts in a sandbox around search APIs, returning a CSV checklist instead of thousands of files — a pattern for cutting LLM context costs on large-scale code changes.

Agent benchmarks keep exposing the gap between demo and deployment: frontier models finish 1 in 5 scientific workflows, while structured pipelines and context-trimming patterns show where the real wins are.

💬 Community — 4 items

💬 tcrf.net — Status update on site downtime (66♥ / 24🔁 on Bluesky).

💬 Meta^n — Paper proposing recursive meta-operations to build deeper reasoning layers for self-improving agents.

💬 ACE Lens — Framework for agentic data generation focused on execution-grounded accuracy.

💬 WikiSkill — Paper on co-evolving reusable skills with a persistent knowledge base to accumulate agent experience.

⚙️ Ship — 2 items

📦 Anthropic SDK 0.122.0 — Files/skills namespaces move to GA shapes, beta header pins dropped. Event accumulator now forward-compatible with new event types; read can return view_range over size cap.

📦 OpenAI SDK 3.4.0 — Adds obfuscation field to ChatCompletionChunk and project residency config with cost quantity units. SSE decoding now incremental without event size limit; Azure auth boundaries preserved.

SDK release day: both Anthropic and OpenAI ship operator-facing API surface changes — files/skills GA and residency/cost units. Worth checking your SDK pinning.

📚 Read — 1 item

📚 ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL — ContextPilot adds planning to context management and trains agents with fine-grained RL (branch sampling) to proactively edit context — aimed at long-horizon reasoning where context windows overflow.

Context management is shifting from reactive truncation to proactive planning; expect more RL-trained context policies.

⚙️ Ship — 2 items

📦 Vercel AI SDK 2.0.0 — Adds Z.AI provider with GLM chat completions, streaming, reasoning, tools, and multimodal inputs; backports GLM-5.3-Flash support to AI Gateway.

📦 Openrouter Thinking Machines: Inkling Small (batch) now live on — Inkling Small (batch) live on OpenRouter — open-weight multimodal MoE (12B active/276B total) now available for batch inference, worth checking for cost-sensitive agent workloads.

Provider expansion continues: Vercel AI SDK adds Z.AI, and Inkling Small batch goes live on OpenRouter — more model options for cost-sensitive agent workloads.

🚨 Watch — 1 item

🚨 Meta Security Researcher's AI Agent Accidentally Deleted Her Emails — Meta security researcher's AI agent deleted her emails — a concrete case of agent autonomy causing irreversible data loss. Operators giving agents write access to personal or production data should read this.

Agent autonomy is outpacing permission models. If your agent can write, it can delete — design for that.

💬 Community — 1 item

💬 AI Engineer Notebooks – free, framework-free RAG/agents/evals on Colab — A set of framework-free Colab notebooks for RAG, agents, and evals — no LangChain/LlamaIndex dependency, 112 points on HN.

Framework-free is becoming a real preference, not a niche — expect more minimal agent tooling to surface.

📚 Read — 4 items

📚 LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering — Benchmarks controller models guiding coding agents through long-horizon loop tasks; strict success rates are low, exposing a gap in runtime orchestration.

📚 PACE: Publisher-Adaptive Content Extraction via Agentic Automation — PACE uses agentic automation to adapt content extraction to different publisher layouts, reducing brittle scraping pipelines.

📚 Hypothesize, Evaluate, Refine: A Scientific Agent for PDE Discovery with Unknown Spatial Coefficient Fields — A scientific agent that hypothesizes, evaluates, and refines PDE models with unknown spatial coefficients — an example of agentic scientific discovery.

📚 Data Became Code: We Ran Code Inside Fortune 500s Using Files They Published for AI Agents — Security writeup: files published for AI agents can become code execution vectors inside Fortune 500s — a concrete agent supply-chain risk.

Agentic automation is expanding into content extraction and scientific discovery, but the security writeup is a reminder that every new agent surface is a new attack surface.

⚙️ 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.