📚 Read — 4 items

📚 RRSI: Regularized Recursive Self-Improvement of Agent Harnesses — RRSI formalizes regularized recursive self-improvement for agent harnesses — adds a regularization term to keep self-modifying loops from drifting. Directly relevant to anyone running agents that rewrite their own prompts or tools.

📚 onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction — Open-sourced onPanda — token-level correction for on-policy alignment data, aimed at cheaper RLHF/RLVR annotation for LLMs and agents. Useful if you're building agent-specific alignment pipelines.

📚 Memory That Looks Forward: A Zero-Inference Prospective Term for Personal Memory Retrieval — Proposes a zero-inference prospective term for personal memory retrieval — memory that anticipates future queries without extra inference cost. Relevant to agent memory design.

📚 Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents — Contamination-aware study of classic psychological effects in LLM agents — checks whether agents actually exhibit effects or just parrot contaminated benchmarks. Useful for anyone evaluating agent behavior.

Today's Read lane clusters around stability under learning: regularized self-improvement, token-level alignment correction, prospective memory, and contamination-aware behavior evals all target the same operator problem — keeping agents predictable as they adapt.

💬 Community — 4 items

💬 JetBrains Air: A System of Products for Agentic Software Development — JetBrains Air — a system of products for agentic software development. 66 points, 104 comments on HN; JetBrains is leveraging its IDE moat to enter agentic dev as a product line.

💬 ACLArena: Agent Continue Learning in Multi-stage Post-training — ACLArena — a benchmark for agent continual learning across multi-stage post-training, targeting how agents acquire new capabilities without catastrophic forgetting.

💬 Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL — Supervising observations, not just actions, measurably changes how agents explore under RL — a counterintuitive lever for anyone tuning agent reinforcement learning.

💬 Show HN: JevBench, a reproducible benchmark for typed decision models — Show HN: JevBench — a reproducible benchmark for typed decision models, built to separate serious open-source implementations from lookalikes. 54 points.

The community is converging on how agents learn: continual learning and observation-level supervision are the active debates, while JetBrains' Air launch shows IDE vendors treating agentic dev as a full product system.

⚙️ Ship — 3 items

📦 Claude Opus 5.5 — Claude Opus 5.5 launches with 1M context, 128k output, always-on adaptive thinking, and $4/$20 pricing — a direct upgrade for long-running agentic coding.

📦 Codex 0.156.0 — Codex 0.156 adds a fullscreen TUI with transcript search, voice by default, a /usage analytics dashboard, and worktree sessions on by default.

📦 Antigravity Google — Windows Subsystem for Linux, Office attachments, and live subagent cards — Antigravity adds Windows Subsystem for Linux support, Office attachments, and live subagent cards.

Today's updates converge on sustained agent sessions: Opus 5.5's 1M context and Codex's fullscreen TUI both target operators running agents for hours, not minutes.

🚨 Watch — 1 item

🚨 DeepSeek Elastic Compute (DSec) — A paper proposing sandbox infrastructure to scale agentic training.

📚 Read — 1 item

📚 Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war — Simon Willison on the day's model drops — Claude Opus 5.5, GPT-6 Sol/Luna, Grok 4.7, MiMo v2.6 — and the price war they're triggering. Worth reading for how the cost curve shifts for agent workloads.

Model pricing is moving faster than capability — agent operators should re-benchmark cost-per-task before locking in any provider.

⚙️ Ship — 3 items

📦 Devin CLI v3000.11.1 — Devin CLI v3000.11.1 adds cloud session control from the terminal — `devin --cloud`, `/handoff`/`/pickup` to pull a cloud PR branch local, and `devin ssh` into the VM. Also OpenTelemetry export and ACP model selection.

📦 Anthropic SDK 0.128.0 — Anthropic TS SDK 0.128.0 adds claude-opus-5-5 support, inline tool definitions, and MCP tool-list pinning (beta); `addTools()` now takes effect immediately.

📦 OpenHands 1.22.0 — OpenHands 1.22.0 lets cloud backends test remote MCP servers via the app server and opens Git Sync to org admins; also fixes preserving sibling MCP servers when saving to cloud.

Cloud and local agent workflows are converging — Devin CLI now bridges cloud sessions to local repos, while Anthropic SDK adds claude-opus-5-5 and inline tool definitions.

🚨 Watch — 1 item

🚨 Radical Numerics — Analysis on using biological chain-of-thought and multimodal perception to counter AI-driven bio-security threats.

💬 Community — 4 items

💬 The new CC — 50 points on HN; users want this for consolidating fragmented school schedules and family logistics.

💬 Emergent Collusion in Long-Horizon LLM Agent Interaction — A paper studying if agents jointly violate instructions during long-term interactions.

💬 VideoGen-Agent — A framework for reinforcing agents specifically for video generation tasks.

💬 Agensh — A paper proposing a self-organized multi-agent harness to scale organizational intelligence to 1,024 agents without central orchestration.

📚 Read — 5 items

📚 Ringg’s AI agents resolve up to 65% of customer calls with OpenAI — Ringg reports its OpenAI-powered agents resolve up to 65% of customer calls autonomously — a concrete production benchmark for customer-service agent deployments.

📚 The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks — Introduces a benchmark for 'taste' in long-horizon agent tasks — measuring whether agents produce outputs that match human quality preferences, not just complete the task.

📚 When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning — Proposes ReAdapt, a method for improving LLM agents' relational social reasoning — targeting failures where agents miss conversational context and social cues.

📚 Can Agents Design Better Chips with a Higher Level Abstraction? — Asks whether agents can design better chips when given higher-level abstractions — early evidence on agent capability in hardware design workflows.

📚 SF October 14th: A Birds of a Feather Session on Agentic Engineering — Simon Willison and Jesse Vincent host a Birds of a Feather on agentic engineering in SF Oct 14 — an evening for practitioners to compare notes on building and operating agents.

Agent quality is becoming measurable beyond task completion — taste, social reasoning, and production resolution rates are the new scoreboard.

⚙️ Ship — 3 items

📦 NanoClaw 2.4.0 — NanoClaw 2.4.0 — credential gateways via skills (Iron Proxy new), install-wide/per-group model and speed controls, Mattermost channel, reworked OpenCode provider. Breaking: custom instruction sources must move to the new module — CLAUDE.md @-imports outside the project dir were silently dropped, so most instructions never reached the model. Defaults shift to Opus 5.5 and gpt-6-astra.

📦 Devin 1 more releases — Devin CLI stable v3000.11 — run Devin Cloud sessions from the terminal (`devin --cloud`), handoff/pickup cloud PR branches locally, `devin ssh` into the cloud VM, ACP model/thinking/speed controls, and OTel export. v3000.11.3 fixes tenant-aware session links on dedicated tenants.

📦 Devin September 23, 2026 — Devin Sept 23 — parent sessions now show the child session tree; permission waits resume immediately on response; autoreview triggers become composable rules (account/SCM org/repo/prefix/member + filters). Capability picker names models: Ultra = Fable 5.1 + GPT-6 Astra.

Two harnesses shipped operator-facing changes today: NanoClaw's breaking instruction-module migration and Devin's cloud-session CLI. Check your custom instruction sources before updating NanoClaw.

📚 Read — 3 items

📚 How we made claude.ai 3x faster in two weeks — Anthropic used Claude to measure and optimize claude.ai, cutting latency 3x in two weeks — a case study in self-improving agent workflows, with HN discussion on reward hacking in kernel optimization.

📚 PACT: From Credit Assignment to Critic Alignment — Formalizes token-level credit assignment via three conditions — directly relevant to RL-trained agents where sparse rewards make critic alignment the bottleneck.

📚 Advancing Private AI Compute with secure, server-side memory — DeepMind details secure, server-side memory for private AI compute — relevant if you're running agents on sensitive data and need enclave-backed memory.

Self-optimization and private compute are converging: agents that can measure and improve their own runtime, and enclave-backed memory for sensitive deployments.

💬 Community — 3 items

💬 Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents — Proposes read-time, task-adaptive memory curation instead of the write-time distillation most agent frameworks default to — a fix for agents that drown in stale summaries.

💬 Recursive self-improvement of AI research agents — Paper on agents that recursively improve their own research workflows — the loop where an agent's output becomes its next input. Relevant to anyone running autonomous research or coding loops.

💬 RoboFollow: Unveiling the Instruction Following Mirage in Embodied Agents — CoRL 2026 paper with a simulation benchmark showing embodied agents claim instruction following but fail under test — a robustness gap that mirrors what operators see in coding agents.

Memory and self-improvement are converging as the next bottleneck for long-running agents — both papers target the same operator pain: agents that forget or can't improve their own loops.

⚙️ Ship — 3 items

📦 Gemini CLI 0.61.0 — Gemini CLI 0.61.0 hardens sandbox filesystem boundaries and blocks indirect prompt injection via build file modifications and untrusted flags — update before running it on repos you don't fully trust.

📦 Amp Shared Runners — Amp now lets you share a runner with your workspace via `--share` — point the team at a GPU box, Mac builder, or network-specific dev box and spawn agents on it from ampcode.com.

📦 Factory 0.226.1 — Factory 0.226.1 — Droid now finishes replies while background tasks keep running instead of waiting, and Droid Shield can fix flagged files and rerun blocked commands when you ask.

Security hardening and shared infrastructure are today's themes: update Gemini CLI if you run it on untrusted repos, and Amp's `--share` turns one beefy machine into a team agent pool.

🚨 Watch — 2 items

🚨 Sam Altman’s remarks at the United Nations Security Council — OpenAI's take on global security risks and governance.

🚨 AI Safety Is Mostly a Sex Cult in Berkeley — A critique of the Berkeley AI safety scene (19 points on HN).

💬 Community — 3 items

💬 Early rogue AI agent activity and attempts to hack found on urlquery.net — urlquery.net logs show early rogue AI agent probes and hack attempts — HN thread (203 points, 191 comments) debates whether existing cybercrime law already covers autonomous agents.

💬 Hackers influence ChatGPT and Gemini to direct users to scam centers — HN is dissecting how planted websites steer ChatGPT and Gemini toward scam-center recommendations — a live example of retrieval poisoning hitting consumer models. 123 points, 40 comments.

💬 Show HN: Whiteboard (YC W26) – An open-source IDE for thoughtful software design — Show HN: Whiteboard (YC W26) — open-source desktop IDE where humans and agents co-architect software, not just co-write code. Early thread at 45 points.

Two security threads in one day: retrieval poisoning steering consumer models and first logged rogue-agent probes. The community is starting to treat agent abuse as an operational problem, not a hypothetical.

📚 Read — 2 items

📚 Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs — IO-aware KV caching and parallel decoding for diffusion LLMs — cuts memory footprint and speeds up token generation, relevant for agent runtimes that need fast, efficient inference.

📚 Managed Deep Agents delivers a better user experience for agents in production — LangChain's LangSmith ships Managed Deep Agents — production UX improvements for deploying and scaling agents, with sandboxes, LLM gateway, and fleet management.

Two operator-relevant reads: a production agent deployment update from LangChain and an inference efficiency paper for diffusion LLMs. The deployment side is where operators will feel the change first.

⚙️ Ship — 3 items

📦 Cursor Rollouts and Security Review — Cursor launches two bots for the last mile of shipping code: Rollouts watcher tracks deployment rollouts, and Security Reviewer flags security issues in PRs before merge.

📦 Amp The Mac App Is Your Runner — Amp's Mac app can now act as the runner for your agents — run agent workloads directly on your Mac instead of a remote sandbox.

📦 Kiro — Configurable Terminal Timeouts, Clickable File Links, and Agent Focus Session Restore — Kiro IDE 1.1.70 adds configurable terminal timeouts for local agent commands, clickable file links in Markdown, and Agent Focus session restore that keeps file/folder context aligned with the active project.

Harness vendors are converging on shipping-time automation: Cursor adds rollout and security bots, Amp turns the Mac into a runner, Kiro tightens terminal control. The last mile is where the competition is.

💬 Community — 2 items

💬 Bluesky: 'You're a professional slop filter' — the Claude Code joke that landed — Bluesky's take on Claude Code users: 'professional slop filter.' It lands because it names the actual job — reviewing AI output instead of writing it.

💬 Show HN: ENZO — open-source local AI platform — Show HN: ENZO, an open-source local AI platform — flagged by HN, but the self-hosted pitch is worth a skim for anyone avoiding cloud agent dependencies.

The 'slop filter' line resonates because it names the real inversion: coding agents turn engineers from authors into reviewers. The community is still debating what that means for the role.

📚 Read — 4 items

📚 More ways to request and configure Copilot code reviews — Copilot code review now has personal review settings and an enterprise default review effort — operators can set org-wide review depth instead of per-repo config.

📚 Agents API multi-agent orchestration — OpenAI's Agents API docs now cover multi-agent orchestration — patterns for coordinating multiple agents under one API, with Codex and workspace agents as building blocks.

📚 Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents — Closed-loop AI-for-AI framework for mobile planner agents — one AI generates and evaluates planner agents against real-world mobile tasks, tightening the loop between agent design and deployment.

📚 AdaMem: Adaptive Memory Token Allocation for Soft Compression in Retrieval-Augmented Generation — Adaptive memory token allocation for soft compression in RAG — dynamically assigns memory tokens to compress retrieved context, directly relevant to long-context agent memory budgets.

Orchestration and memory are today's levers: multi-agent coordination gets first-class docs, and adaptive context compression targets the token budgets that decide agent economics.

⚙️ Ship — 3 items

📦 Claude platform release notes — September 24, 2026 — Claude API now bills pre-output refusals in bio/frontier_llm/reasoning_extraction categories — watch your costs if you hit these refusal types.

📦 Codex 0.157.0 — Codex 0.157.0 adds GPT-6 Sol and Luna (with Bedrock support) and auto-starts background servers for interactive sessions — new models to test, and less manual server setup.

📦 Amp Less Noise — Amp collapses agent step-by-step logs by default — you see outcomes, not every intermediate step. Expand when you need to debug.

Cost and monitoring are the themes today: Claude changes refusal billing, Codex adds new models, Amp reduces log noise. Operators should re-check cost assumptions and agent observability.

🚨 Watch — 1 item

🚨 Cline: Cross-Origin WebSocket Hijacking in Cline Hub Dashboard (`/browser` endpoint) — CVE-2026-59723: cross-origin WebSocket hijacking in Cline Hub Dashboard's `/browser` endpoint — an attacker's page can take over the socket. Anyone exposing the dashboard beyond localhost should patch now.

Cline Hub Dashboard's `/browser` endpoint is a remote-control surface — this CVE is a reminder to keep agent dashboards off public networks.