⚙️ Ship — 2 items

📦 Claude API: on-demand compaction beta + Managed Agents auto permissions — Messages API can compact a conversation on demand (beta `compact-2026-09-04`); Managed Agents permission policies add `auto` so the server evaluates each tool call.

📦 Anthropic SDK 1.6.0 — SDK 1.6.0 exposes the compaction parameter and signed blocks (beta), auto-mode Managed Agents permissions, and async credential token providers.

Compaction on demand is the operator-relevant move: you control when to summarize and can keep recent turns verbatim, which matters for long-running agent sessions.

💬 Community — 5 items

💬 Mistral X Mozilla: Private, Multilingual AI Browsing — Mistral and Mozilla partner on private multilingual AI browsing — HN's 400-point thread centers on local vs cloud inference for privacy.

💬 Apple's Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows — Code shows Apple's Siri AI can be swapped for Claude or ChatGPT — HN's 225-point thread debates Apple's closed ecosystem vs open model choice.

💬 Show HN: Pelican-bicycle alternatives (updated for 2026) — HN user tests 30 LLMs on SVG generation prompts inspired by Willison's pelican benchmark — updated for 2026 models.

💬 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation — Benchmark Radar is a searchable living database of AI benchmarks — aggregates sources, score histories, and discovery for eval selection.

💬 ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement — ModularRSI proposes modular recursive self-improvement for agent harnesses — generalizable across tasks, not task-specific.

Privacy and model choice are the live debates this cycle — not raw capability.

📚 Read — 5 items

📚 Coding agents usually can't price on outcomes. Ours can. — Agentic Batch Changes prices on outcomes — you only pay for changesets that merge. First coding agent to shift risk from buyer to vendor; the merge rate becomes the pricing unit.

📚 Introducing Agentic Batch Changes: the frontier agent for code change at scale — Agentic Batch Changes runs codebase-wide changes across hundreds or thousands of repositories — org-wide refactors as one agentic workflow rather than a script. The scale claim is the story.

📚 Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States — Trains LMs to simulate user mental states before responding — 'understand the person behind the prompt' as a trainable objective, not a prompt-engineering trick. Relevant to agent UX and goal inference over literal instruction-following.

📚 Mistral and Mozilla are bringing open, private and multilingual AI to your web browser — Mistral and Mozilla partner to put open, private, multilingual models in the browser — local inference as the privacy answer. Worth tracking if on-device agents become a real deployment target.

📚 GitHub Copilot suggests custom properties definitions — Copilot now suggests allowed values when creating repo custom properties — org admins managing Copilot policies get autocomplete on their governance surface. Minor, but Copilot's admin tooling is getting smarter.

Coding-agent pricing is starting to move from per-seat and per-token to per-outcome. Sourcegraph's merged-changeset pricing is the clearest signal yet — the rest of the market either follows or explains why not.

💬 Community — 1 item

💬 Show HN: Loss — a tiny satire about AI progress — HN satire on agent sprawl: you start as a manual inference unit pressing a button per token, then agents, sub-agents, approvals, swarms take over. Universal Paperclips for the agentic era.

When the top community post is a parody of orchestration layers multiplying faster than anyone can evaluate them, it's worth asking which of your own approval layers actually earn their keep.

⚙️ Ship — 2 items

📦 Devin CLI updated — Devin CLI v3000.10.31 fixes permission precedence: `Exec(rm)` no longer blocks when a broader `ask`/`allow` rule covers the tool; `Exec(*)` now matches every command. Also new this week: GPT-6 Astra batches shell commands, `/code` `/smart` `/bypass` mode commands, and `disabled_tools` config.

📦 OpenHands 1.19.0 — OpenHands 1.19.0 — agent profiles can now be scoped to specific MCP servers; adds GPT-6 Astra support. Fixes credential injection in static-server (HTML-escape + no-store) and Azure DevOps SSH remote recognition.

Permission and MCP scoping are the themes today — both Devin and OpenHands are tightening control surfaces for operators running agents in shared environments.

🚨 Watch — 1 item

🚨 Artificial intelligence and biosecurity: capabilities, threat pathways, and defense-in-depth governance — A paper proposing governance frameworks to mitigate AI-driven biological threats.

📚 Read — 1 item

📚 ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments — Proposes turning scientific codebases into environments agents can learn from — a potential source of training and eval signal for code agents.

Code-as-environment is becoming a research direction; expect more papers on turning real codebases into agent training grounds.

⚙️ Ship — 2 items

📦 Claude API: on-demand compaction (beta) + Compliance transcripts — Claude API adds on-demand conversation compaction (beta) — send a `compaction` parameter, get a signed summary block, keep recent turns verbatim; Compliance API now returns Claude in Chrome session transcripts for Enterprise.

📦 ACP 1.22.0 — ACP 1.22.0 stabilizes tool call name and adds session notices schema — protocol-level changes for agent-client interop.

Compaction is becoming an operator-controlled primitive, not a runtime decision — worth testing before context costs bite.

💬 Community — 2 items

💬 Show HN: Pizza Bot – An inbox for AI agents that work in the background — Self-hosted desktop app (Mac/Win/Linux) that runs AI agents in the background and surfaces their output in an inbox — 59 points and 33 comments of HN debate on background-agent ergonomics.

💬 Open-sourced jev architecture last year with model,paper and dataset — HN thread on a non-autoregressive architecture for JSON-schema-constrained probability prediction, open-sourced with model, paper, and dataset. Directly relevant to structured outputs in agent tool calling.

Both threads circle the same gap: agents that run unattended need better ways to produce and deliver structured output. Background-agent UX and constrained decoding are two sides of the same problem.

📚 Read — 4 items

📚 Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents — XConf proposes confidence estimation grounded in experiential outcomes rather than model introspection — a calibration approach that could tell agents when to trust their own reasoning before acting.

📚 HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses — HypoEvolve applies genetic selection across multi-agent LLM outputs to evolve scientific hypotheses — an orchestration pattern worth stealing for any agent swarm doing open-ended search.

📚 GraphEcho: Structural Redundancy and Evidence Provenance in LLM Graph Agents — GraphEcho dissects where LLM graph agents duplicate structure and lose track of evidence provenance — useful for anyone building retrieval or knowledge-graph agent loops.

📚 ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search — ZGCM-1 is a fully open model tuned for math and agentic search, claiming extreme efficiency — worth a look if your agents spend tokens on search-heavy reasoning.

Confidence and provenance are the through-line today: three of four papers attack how agents know what they know before acting.

💬 Community — 2 items

💬 @freyjavalentine.bsky.social: Saw a term calling the act of shitting out Claude code without review and expecting your coworkers to fix it — Bluesky thread coining a term for unreviewed Claude code dumped on coworkers — the 'agent slop' debate continues.

💬 Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS — Show HN: Aclif — agent CLI framework with a single grammar and canonical names across SaaS providers. Ships provider schemas in the binary, no credentials needed for discovery.

The agent CLI space is consolidating around canonical schemas, while the 'unreviewed Claude code' problem finally gets a name.

⚙️ Ship — 2 items

📦 Codex 0.155.0 — Codex 0.155 adds experimental /voice conversations, task archiving/deletion, Touch ID verification for MCP requests, and configurable daemon update schedules.

📦 OpenHands 1.20.0 — OpenHands 1.20 lets agent profiles select which secrets are available and automations pick saved profiles; Docker conversation runtime settings are now forwarded.

Both releases tighten operator control: Codex adds Touch ID for MCP and configurable daemon updates, OpenHands scopes secrets per profile. The trend is toward more granular session management.

📚 Read — 2 items

📚 SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness — Proposes recursively scaling auto-research loops for agent harnesses — targeting unattended coding agents that move beyond supervised code completion.

📚 RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning — Self-retiring on-policy distillation for agentic RL — the distillation schedule adapts and stops when the student stops benefiting, instead of following a fixed plan.

Both papers replace fixed training schedules with adaptive control — agent training is converging on self-regulating loops.

⚙️ Ship — 2 items

📦 Anthropic Life Sciences Verification Program — Anthropic opens Mythos, Opus, and Sonnet to life science professionals with more permissive biology safeguards — relevant for biotech agent builders needing domain-specific model access.

📦 ACP 1.23.0 — ACP 1.23.0 adds a v1 session notice capability (unstable) — a new protocol signal for session lifecycle, useful for client/server implementations tracking session state.

ACP keeps adding session lifecycle primitives; Anthropic's life sciences program shows model access expanding into regulated domains — both worth tracking for domain-specific agent builders.

💬 Community — 1 item

💬 Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents — YC W26 launch: Skillsync makes AI chat sessions portable across coding agents — move context between Claude Code, Cursor, etc. 61 points, 53 comments on HN; the debate is whether cross-agent portability is a real workflow need or a solution looking for a problem.

Cross-agent session portability is becoming a product category — operators juggling multiple coding agents may soon have a standard way to move context.

📚 Read — 3 items

📚 Closed-World Resolution Against Tool Hallucination in LLM Agents — Proposes closed-world resolution to stop LLM agents from hallucinating tool calls — a top production failure mode when agents invoke tools that don't exist or misuse schemas.

📚 MAGS: Multi-agent Auto-formalization Guarantees Safety for Agentic Outputs — Multi-agent auto-formalization claims formal safety guarantees for agentic outputs — worth a look if you need verifiable agent behavior beyond prompt-level guardrails.

📚 Do AI Agents Understand Computer Architecture? — Benchmark probing whether AI agents actually understand computer architecture — useful signal for coding-agent evals and where they break on low-level reasoning.

Reliability research is converging on the failure modes that actually bite in production: hallucinated tool calls and unverifiable outputs.

💬 Community — 1 item

💬 2,151 IPs banned from The Cutting Room Floor wiki for Claude Code scraping — 2,151 unique IPs hit The Cutting Room Floor wiki with Claude Code and got banned — sites are actively blocking agent traffic. Worth noting for anyone running agents against public wikis.

Agent traffic is now a first-class concern for site operators — expect more aggressive blocking as scraping agents proliferate.

⚙️ Ship — 2 items

📦 Antigravity Google — Custom agent controls and keyboard navigation — Antigravity adds custom agent controls and keyboard navigation — more direct steering of agent behavior in the harness.

📦 xAI API release notes — September — xAI's September API release notes are live — Grok 4.6 is the current model line. Worth a scan for tool-use and structured-output changes that affect agent builds.

Control surface is the theme today — Antigravity adds direct agent steering while xAI refreshes its API line. Ergonomics, not raw capability, is where the moves are.

📚 Read — 4 items

📚 Characterizing Web Search by Conversational LLM Agents: From Search Decisions and Strategies to Results and Responses — Empirical study of how conversational LLM agents decide when to search, which strategies they pick, and what they return — useful signal for tuning your own agent's search policy.

📚 Agentgit: a Git host for AI agents with no accounts — Git host where the repo name is the identity — no accounts, tokens, or keys; agents push to a name and hand off the URL. Open source and self-hostable.

📚 HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness — Training-free embodied navigation via a harness that unifies VLN methods — no fine-tuning, the harness coordinates perception and action. Relevant if you're building embodied agents.

📚 AI for Games in the Foundation Model Era — Survey of foundation-model game agents — where LLM-driven NPCs, planners, and world models stand, and what's still missing.

Agent infrastructure is splitting into two tracks: empirical studies of how agents actually use tools, and frictionless primitives for agent-to-agent handoffs.

⚙️ Ship — 3 items

📦 Anthropic SDK 0.127.0 — Adds `compactBeforeNextTurn()` to the tool runner for explicit context compaction, and stops retrying streamed request bodies. Worth pulling if you run long-context agent loops.

📦 Antigravity Google — New Permissions System & Terminal and Git version control for Enterprise — Antigravity ships a new permissions system plus terminal and Git version control for Enterprise. Permission changes are the kind of thing operators need to review before rollout.

📦 OpenClaw release-publish/ee1e5f21fd22 — New cut from the core runtime. Release notes weren't captured this cycle — check the tag for operator-facing runtime changes.

Compaction control and permission systems both shipped today — the operational surface of agents is getting more explicit knobs, not just more capability.

🚨 Watch — 1 item

🚨 Gemini — Google's AI breached three companies in a security test; proves agentic capabilities outpace current guardrails.