📚 Read — 4 items

📚 MemT-1.0: Towards Memory-Native Evolution — Proposes a memory-native architecture that embeds continuous memory directly into foundation-model computation to improve long-context efficiency — relevant to any agent stack that leans on memory for long runs.

📚 Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides — arXiv cs.AI paper (submitted 27 Jul 2026) proposing typed MCP overrides as a mechanism for persistent human control over LLM agents — a concrete pattern for operators wiring authority and oversight into MCP-based deployments.

📚 Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents — Targets the failure mode where small language model agents get stuck in repetition loops — a distillation method that detects and penalizes stagnation during training to improve reasoning efficiency.

📚 CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents — Synthesizes executable training environments for long-horizon AI coworkers using workflow-guided task construction and programmatic validation — takes on the 'how do we build training environments for AI coworkers' problem.

MemT-1.0, Progress-Aware Distillation, and CoworkEnv all push on agents staying effective over long runs, at three different layers: memory architecture, training objective, and training environment.

💬 Community — 1 item

💬 DeepSeek banned from U.S. federal civilian agencies under new DHS order — HN discussion of the new DHS order banning DeepSeek from U.S. federal civilian agencies — 549 points and 501 comments, with the Federal Register document (2026-21447) linked in the thread.

⚙️ Ship — 2 items

📦 Gemini CLI 0.65.0 — Fixes deletion of resumed session history on quick exit (#29198) and duplicate tool response turns when resuming sessions; also enforces read-only workspace settings in untrusted folders, prevents infinite OAuth verification and retry loops, and adds support for custom OTLP headers in telemetry configuration.

📦 Kiro 1.4.10 — Memory Setting Defaults to Off, Workflow Plan Feedback, and Resilient Powers — The Memory setting's default changes to Off, approved workflow steps follow your plan feedback, and Powers keep working when a server is unavailable; several chat fixes also land.

Both releases change default agent behavior: Gemini CLI 0.65.0 enforces read-only workspace settings in untrusted folders, and Kiro 1.4.10 changes the Memory setting default to Off.

📚 Read — 2 items

📚 ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model — IBM Research's ALTK-Evolve tops the τ²-bench leaderboard on customer service tasks with learned memory and no frontier model — relevant reading for anyone designing agent memory.

📚 JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs — Delegates agent tool choices to small decision models via fitted confidence thresholds, cutting cost and latency — a benchmark and deployment harness for decision models that share control with LLMs.

Both items show agent results without leaning on frontier-scale models — ALTK-Evolve tops the τ²-bench leaderboard on customer service tasks via learned memory, while JevAgentHarness cuts cost and latency by handing tool choices to small decision models.

⚙️ Ship — 1 item

📦 ACP 1.26.0 — Adds a client system prompt schema (#2361), removes next edit suggestions (#2266), and clarifies that a forked session is active (#2359) — all three changes are unstable-spec only.

💬 Community — 2 items

💬 Patch the Planet: global response coordination as agents uncover vulnerabilities — HN thread on OpenAI's Patch the Planet — coordinating global response as agents uncover vulnerabilities. 160 points / 95 comments; commenters are reacting to the released skills, including one that checks PHP versions.

💬 Using Claude to reduce the attack surface of code — HN thread on Anthropic's piece about using Claude to reduce the attack surface of code. 116 points / 76 comments; discussion pulls in kernel-vulnerability cleanup experience.

Both kept threads sit at the same intersection: agents applied to code security — one on coordinating global response as agents uncover vulnerabilities, the other on using Claude to shrink code attack surface.

📚 Read — 3 items

📚 MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records — MMEnterprise-Bench evaluates multimodal agents on reasoning over evolving workplace records with conflicts and missing evidence.

📚 Using thousands of AI agents to uncover new leads in biological data — OpenAI's writeup on using thousands of AI agents to uncover new leads in biological data.

📚 Benefit Street Partners answers questions faster with Codex — OpenAI customer story on Benefit Street Partners using Codex to answer questions faster.

🚨 Watch — 2 items

🚨 Degraded performance for Claude Fable 5 — Elevated errors on requests to Claude Fable 5, under investigation since 19:49 UTC on Oct 21. Affects claude.ai, the Claude API (api.anthropic.com), Claude Code, and Claude Cowork.

🚨 OpenClaw: Browser CDP discovery could accept blocked WebSocket URLs (CVE-2026-62197) — CVE-2026-62197, medium severity: OpenClaw's browser CDP discovery could accept WebSocket URLs that were supposed to be blocked — URL restrictions in browser sessions could be bypassed.

💬 Community — 2 items

💬 Claude Duet — 289 points and 245 comments on HN for the Claude Duet announcement on claude.com; the opening comment pushes on the waitlist — why signup isn't requestable from an existing signed-in Claude session.

💬 OpenAI models are coming to Google Cloud — 220 points and 94 comments on HN for OpenAI models coming to Google Cloud; the thread opens on the 'Microsoft remains OpenAI's primary cloud partner' disclaimer at the bottom of the announcement.

Both kept threads are HN discussions of AI-lab rollout news, and in both the opening comments target the fine print — the Claude Duet waitlist (why it isn't requestable from an already-signed-in Claude session) and the 'Microsoft remains OpenAI's primary cloud partner' disclaimer on the Google Cloud announcement.

⚙️ Ship — 2 items

📦 Claude platform release notes — October 21, 2026 — API version 2026-10-21 is released, with new versions of the client SDKs and ant CLI that send it by default — Python SDK 2.0.0, TypeScript SDK 1.0.0, Go SDK 2.0.0, and ant CLI 2.0.0 among them. Those versions carry breaking changes that don't depend on the API version and are licensed under Apache 2.0 instead of MIT; user profiles also add enforcement_status and enforcement_actions fields.

📦 Pi 1.4.0 — Codemode parallel scripts and in-process subagents now queue for a free worker instead of failing (PI_CODEMODE_MAX_WORKERS raises the limit), and compaction and branch summaries use the model's output limit so long reasoning no longer makes them fail. Also: codemode models.classify() accepts images for TypeSafe's Jev and Cloudflare's Clef classifier models, and a models.json configuration reference is generated from the schema.

📚 Read — 2 items

📚 Claude Mods: Flipping the Harness Inside Out — Dan McAteer spent weeks using Claude Mods to mold Claude Code to his AI engineering workflow; the piece frames Mods as an early form of the 'Attention Interface' — flipping the harness paradigm from facing the machine to facing the human.

📚 AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics — Trains the planner model of a robot harness via supervised fine-tuning and multi-turn reinforcement learning across diverse harnesses — coding-agent post-training methods applied to agentic robotics.

Both pieces put the harness at the center: Claude Mods flips it to face the human, while AgenticRoboRL brings coding-agent post-training — supervised fine-tuning plus multi-turn RL — to planner models inside robot harnesses.