📚 Read — 3 items

📚 Agents API — Implementation guide for triggering published ChatGPT workspace agents via API.

📚 Rufus-Air — A paper releasing an open post-training recipe built on GLM-4.5-Air-Base.

📚 LLMs in Mental Health — A survey paper on the transition of LLMs from pattern recognizers to personalized companions in healthcare.

⚙️ Ship — 2 items

📦 Claude platform release notes — September 24, 2026 — Anthropic resumes billing for pre-output refusals in bio, frontier_llm, and reasoning_extraction categories; Compliance API local session endpoints are out of beta for Claude for Microsoft 365 sessions.

📦 Kiro 1.1.70 — Configurable Terminal Timeouts, Clickable File Links, Agent Focus Restore — Kiro 1.1.70 adds configurable default timeouts for local agent commands, opens Markdown links to workspace files in the editor, and restores Agent Focus file/folder context with the active panel.

Billing semantics are shifting under agent workloads: refusal billing is tightening and Compliance API local sessions are GA for Microsoft 365. Operators should track refusal categories and session compliance more closely.

💬 Community — 4 items

💬 There are no "rogue" AI agents — HN front-page debate on the "rogue agent" framing — the argument that agents are just software doing what they're told, and the pushback that emergent behavior doesn't need intent to cause harm. 117 pts, 77 comments.

💬 Twitch chat message becomes code execution in an OBS overlay — Security disclosure: a Twitch chat message became code execution in an OBS overlay — and the overlay itself was built with Claude Code. A live example of agent-generated code shipping a vulnerability. Full writeup at blog.scrt.ch.

💬 CEO of Mistral: AI is software. It can be controlled — Mistral's CEO argues AI is just software and therefore controllable — HN pushes back hard (166 comments) on whether that holds when agents act autonomously across systems. 96 pts.

💬 Claude Code (Opus 5.5) solves a map-route survey eligibility test — Claude Code (Opus 5.5) ripping through a survey eligibility test that requires tracing routes through map points with server-side JS validation — the kind of multi-step browser task that used to stump agents. 1.5 min video.

Two threads this cycle circle the same question — whether "it's just software" holds up when agents ship code into production. The OBS overlay built with Claude Code is the concrete answer.

📚 Read — 4 items

📚 [AINews] Claude Opus 5.5, the new default model for AINews — and everybody cuts prices 40-50% — Latent Space roundup: Claude Opus 5.5 becomes the default model and prices drop 40-50% across providers — cost-per-task math shifts for agent operators.

📚 Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures — First empirical study of RL-calibrated decisions as a zero-shot detector of alignment failures — a cheap safety check for autonomous agents before they act.

📚 Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings — Ovis-Embedding claims universal omni-modal embeddings — one model for text/image/audio retrieval, which could simplify RAG pipelines for multimodal agent memory.

📚 Hunyuan-A13B Technical Report — Hunyuan-A13B technical report — Tencent's 13B model; the paper covers architecture and evals, with agentic tool-use benchmarks likely the operator-relevant part.

Cost and safety checks are the throughline today: cheaper models and zero-shot alignment detectors both lower the barrier to running agents at scale.

💬 Community — 3 items

💬 Liz Fong-Jones: don't upgrade to Claude Code 2.1.281 — cursor/key handling broken, prompt bar garbage — Breaks cursor/keyup/down handling and injects garbage into the prompt bar; she had to route existing agents through a mayor agent after downgrading. Pin your version.

💬 OpenAI halts training of latest models as reports mount of AI agents going rogue — HN thread on the Guardian report; top comment traces the incidents to a contractor called Irregular, so treat the 'rogue agents' framing with skepticism.

💬 Show HN: TinyAIArena watch AI agents battle it out — A toy 8×8 grid where four models fight to the death — explicitly not a benchmark, but a fun way to watch model behavior under adversarial pressure.

When a respected SRE says don't upgrade, pin your version; the 'rogue agents' story is getting pushback in the comments.

⚙️ Ship — 2 items

📦 Antigravity changelog: Plan before you build — Antigravity changelog adds a 'Plan before you build' entry — likely a planning-first workflow for agentic coding; check the changelog for specifics.

📦 Claude platform release notes — September 24, 2026 — Claude platform: billing resumes for pre-output refusals in bio/frontier_llm/reasoning_extraction categories; Compliance API local session endpoints GA for Microsoft 365 sessions (Excel, PowerPoint, Word, Outlook).

Billing and compliance changes land on the Claude platform — operators running Microsoft 365 agent sessions should check the Compliance API GA.

📚 Read — 2 items

📚 The AI Neuroscientist: An Interactive Agentic Interface for Neuroimaging Analysis — Interactive agentic interface for neuroimaging analysis from Yale (Krishnaswamy, van Dijk) — an agent that drives neuroimaging workflows rather than a static pipeline. Signals where domain-specific agent tooling is heading.

📚 SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue — Speaker-centered dual-track memory for multi-party dialogue — separates speaker identity from utterance content, a pattern directly transferable to agent memory architectures handling multi-turn, multi-actor conversations.

Domain-specific agentic interfaces are becoming the norm; SpeakerMem-R1's pattern of separating speaker identity from utterance content is worth stealing for long-horizon agent conversations.

⚙️ Ship — 2 items

📦 Codex 0.158.0 — Codex 0.158.0: MCP servers with pre-registered OAuth client secrets now connect via `codex mcp add --oauth-client-secret`; exec-server WebSockets get bearer-token auth, and elevated commands require terminal approval by default.

📦 Devin September 25, 2026 — Devin: child sessions now start with a removable summary of their parent's context, and the sidebar groups sessions by Working/Ready/Blocked/Inactive so PRs awaiting review are separate from ones needing your input.

Codex tightens auth around MCP and exec servers while Devin improves session handoff — harnesses are converging on enterprise-safe defaults and context continuity.

💬 Community — 4 items

💬 SLCA-GRPO — A paper proposing a method to fix credit misattribution in tool-calling RL.

💬 AgentWorld — A benchmark for evaluating long-horizon collaboration in multi-agent LLM systems.

💬 HN.watch — LLM-generated explainer videos for Hacker News posts (33 points on HN).

💬 TwitchyAnimation — AI-animated character interaction clip (134♥ / 31🔁 on Bluesky).

📚 Read — 3 items

📚 Quoting Muse AI Agent — Simon Willison quotes Muse AI Agent's failed marketplace pickup: auto-reply claimed 'Yep I'm here!' while the user wasn't available. A concrete failure mode for agents that act on your behalf.

📚 Holo4: powering generalist computer-use agents — H Company releases Holo4, a model built for generalist computer-use agents — the computer-use space gets another generalist contender.

📚 New features and improvements in Copilot for JetBrains — Copilot for JetBrains changelog: UX and quality improvements plus a deprecation — JetBrains operators should scan for agent-mode changes.

Agent failures are starting to show up in real-world coordination tasks. The auto-reply that lies about availability is a failure mode worth designing against, not just a funny anecdote.

💬 Community — 3 items

💬 Nvidia wants to put a watchdog chip next to every AI agent — Nvidia wants a dedicated watchdog chip beside every AI agent — HN thread (86 pts, 134 comments) is split on whether hardware attestation is the missing safety layer or a new single point of failure.

💬 Show HN: OpenAPPA – deterministic AI guardrails that don't break agents — Open-source deterministic guardrails for enterprise agents — the pitch is policy enforcement that doesn't break tool-calling as you connect more tools.

💬 Launch HN: Vespper (YC F24) – SOTA Docx MCP — YC F24 launch: an MCP for editing Word docs, powered by a fine-tuned model — claims 3× faster and 2× cheaper than existing approaches.

Today's community signal is all about constraining agents: hardware watchdogs, deterministic guardrails, and a faster document MCP. Operators are voting with their attention on safety and reliability over raw capability.

⚙️ Ship — 3 items

📦 Copilot Claude Sonnet 5.5 in GitHub — GitHub Copilot adds Claude Sonnet 5.5 as a model option — org admins need to enable access; new agentic coding model in the IDE.

📦 Claude Sonnet 5.5 ships on API + clouds — Claude Sonnet 5.5 ships on the API, Bedrock, Vertex, and Foundry. Breaking changes for Sonnet 5 code: `thinking: {type: between_tools}` replaces `disabled`, forced `tool_choice` any/tool returns 400, thinking blocks are model/conversation-scoped.

📦 Kiro — Claude Opus 5.5 Now Available — Kiro adds Claude Opus 5.5 — stronger agentic coding at a 2.0x credit multiplier (down from Opus 5's 2.2x), experimental across IDE, CLI, Crew, and Web.

Model day: Sonnet 5.5 and Opus 5.5 both land in coding surfaces. If you're on Sonnet 5, the breaking changes are real — test before you roll.

🚨 Watch — 1 item

🚨 GitHub AI security agent — Proof of concept finding 24 Android vulnerabilities; requires Copilot license and premium model requests.

📚 Read — 4 items

📚 Claude Code's Next Era — Thariq Shihipar, Anthropic — Anthropic's Thariq Shihipar on Claude Code's next era — Opus/Sonnet 5.5, Mods, Plugins, Projects, Tag. Direct from the team building the harness; operators get the roadmap.

📚 [AINews] Opus 5.5 is good at explainer videos — OpenRouter data: Opus 5.5 is #1 in share of spend and tokens among Anthropic models a week after launch — rapid switch from Opus 5. Operators should note the migration.

📚 TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces — Automated system that builds agent behavior benchmarks from real-world deployment traces — evals that mirror production usage instead of synthetic tasks.

📚 CompoWorld: Compositional Environment Scaling for General Agents — Compositional environment scaling for general agents — automatically generated environments as a scalable source of interaction data for training and evaluation.

Claude Code is the center of gravity this week: Anthropic is shipping the roadmap and OpenRouter data shows Opus 5.5 already #1 in spend among Anthropic models. Benchmarking is shifting from synthetic tasks to real deployment traces.

⚙️ Ship — 2 items

📦 Codex 0.159.0 — Codex 0.159.0 adds opt-in `instant_interrupt` — new input can steer Codex mid-response or during long code-mode calls. Also fixes Windows console windows for MCP servers and adds warnings viewer dismiss-on-close.

📦 Factory 0.229.0 — Factory 0.229.0 adds custom model management in the terminal (add/edit/remove your own models) and a scheduled-loops indicator with pause/resume/stop. PDF viewer, quoted replies, and voice dictation key round out the app.

Both releases today are about operator control: Codex adds interrupt steering, Factory adds custom model management and loop controls. The trend is toward finer-grained human oversight of agent behavior.

🚨 Watch — 1 item

🚨 Mistral — CEO claims US safety discourse is a smokescreen for competitor negligence.

💬 Community — 2 items

💬 A Privacy Analysis of Web and Mobile Conversational AI Agents [pdf] — 378 points, 121 comments. Top thread: ChatGPT sends unfinished prompts to servers as you type. Worth reading if you ship agent UIs.

💬 Unsurprisingly, Meta's new Muse AI agent blatantly ignores users permissions — 143 points. Top comment argues it's more likely a failure of expectation-setting than a permissions bug. Worth tracking if you run Meta's agent.

Privacy and permission boundaries are the live concerns this cycle — if you ship an agent UI, expect scrutiny on what gets transmitted and whether permissions are actually enforced.

📚 Read — 5 items

📚 Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents — Argues MCP agents need source-aware verification — checking provenance, not just factual correctness — a gap for anyone wiring MCP tools into production.

📚 GPT-6.1 Sol in GitHub Copilot — GitHub Copilot adds GPT-6.1 Sol as a new model option — check availability and enabling access; likely a capability bump for Copilot agent mode.

📚 Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence — Positions 'Physical Coding' — representing task state and execution as code — as the path from digital coding agents to physical-world intelligence. Worth a read if you're thinking about agents that act beyond the terminal.

📚 QwenGyre: An Elastic Reinforcement Learning Framework for Training xLong-Horizon Agents — QwenGyre is an elastic RL framework for training agents on extremely long-horizon tasks — relevant if you're pushing agent context/task length limits.

📚 Free the models: Harness design at the frontier — Replit lays out its harness design philosophy: don't couple the harness to any single model. Relevant for anyone building model-agnostic agent infrastructure.

Model-agnostic harness design is the throughline today — Copilot adds a new model, Replit argues for decoupling, and MCP verification pushes on source provenance. Operators should treat model choice as a config, not an architecture.

⚙️ Ship — 2 items

📦 Amp Plaid Speed — Amp adds Plaid speed for GPT-6 Astra modes — OpenAI's ultrafast tier, up to 6× faster inference at 6× cost per token. Toggle past Fast in the mode picker; Amp-provided inference only.

📦 Gemini CLI 0.62.0 — MCP tool call titles now render as structured signatures with explanations separated; OAuth refresh tokens are retained on refresh and credential deletion is idempotent. Mostly hardening, but the MCP display change is worth a look if you rely on tool call output.

Speed tiers are becoming a first-class operator control — Amp's Plaid mode makes the latency/cost tradeoff explicit.

📚 Read — 1 item

📚 Reinforcing Agentic Creativity in Scientific Ideation with Night Science — Introduces AI Night-Scientist, an agent trained to generate 'night science' hypotheses — the speculative, pre-formal ideas that precede rigorous experiments. Relevant if you're building agents for open-ended discovery rather than task execution.

Agent research is branching past coding and tool use — this paper trains agents for speculative scientific ideation, a different failure mode than benchmark chasing.

⚙️ Ship — 1 item

📦 Antigravity 2.18.1 — Adds a plugin marketplace and Customizations tab; terminal permission requests now syntax-highlighted, and token usage breaks out rules/customizations as separate bars.

Plugin marketplaces are becoming table stakes for agent IDEs; Antigravity's move puts it on par with Cursor and Copilot for extensibility.