Coding agents

• Claude Code v2.1.222 fixes worktree-isolated sessions being able to run destructive git against the main checkout, and closes a PreToolUse auto-allow bypass in background agent tasks, security-relevant if you run subagents/worktrees.
• Claude Code v2.1.216 adds sandbox.filesystem.disabled to skip filesystem isolation while keeping network egress control, and fixes a quadratic slowdown in long sessions that was stalling multi-turn resumes.
• Claude Code v2.1.215: /verify and /code-review no longer auto-run, invoke them explicitly now.
• Claude Code v2.1.212: /fork now clones your conversation into a background session instead of an in-session subagent (that's now /subtask), plus claude auto-mode reset.
• GitHub engineering on turning one giant AI-generated PR into a reviewable stack: teach agents to decompose output into ordered stacked PRs instead of one unreviewable diff.
• alibaba/open-code-review: Alibaba's internal AI code-review CLI (deterministic pipelines + LLM agent, line-level comments, NPE/thread-safety/XSS/SQLi ruleset) now open source, Anthropic/OpenAI compatible.
• EveryInc/compound-engineering-plugin: skill pack for Claude Code/Codex/Cursor aimed at making each unit of work compound off the last, moved to a root-native plugin layout.
• uber/ADR: Uber's open-sourced observability/threat-detection layer for enterprise agents (Cursor, Claude Code, Codex, customer-facing bots), paper accepted at MLSys 2026.
• browser-use/video-use: open-source video editing driven entirely from Claude Code chat (cuts filler, color grades, assembles from raw footage).

LLM mechanics & models

• Kimi K3 2.8T-A50B: Moonshot's largest open model release yet, pitched as Opus 4.8-class capability at Sonnet 5 pricing, open-weight frontier gap keeps closing.
• GLM-5.2: Nathan Lambert flags it as a capability step-change for open agents, alongside a new paper on open RL recipes for terminal agents.
• llm-anthropic 0.26: Simon Willison's llm plugin adds claude-fable-5, claude-sonnet-5, claude-opus-5 support, riding the LLM 0.32 feature set below.
• Claude Fable 5 and new AI safety fables: Anthropic's silent classifier-based manipulation of AI-research queries now applies uniformly across safety domains, still contentious on transparency.

Agent memory

• Zero-Mem: proposes agent memory ops that skip extra LLM calls for writing/retrieving memory, cutting recurring token/latency cost; HN pushback is that the real win is preserving original traces for auditability, not the zero-token framing, compression risks silently dropping evidence.
• Why an LLM's memory gets expensive and how to fix it: explainer on why long-context prompts cost disproportionately more than short ones at the same model/hardware, KV-cache mechanics.
• TencentDB Agent Memory: team-level memory hub turning chats/docs/code into four shared asset types (Chat Memory, Skill, LLM-Wiki, Code-Graph) reusable across agents and frameworks.

Agentic SaaS

• Unpacking ChatGPT Work: outside reconstruction of how Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills and Tools actually fit together in ChatGPT Work.
• OpenAI ships education-focused plugins for ChatGPT Work and Codex, guided modes for K-12 teachers and college students/educators.
• SnapState: persistent state layer for AI agent workflows, hackernews launch.
• Phind: GPT-4-powered dev search engine; HN take is the SEO-noise filtering is the real value, but hallucination on thin-result queries and unbounded API cost undercut it.
• Glasp MCP Connector: exposes your Glasp highlights/notes as an MCP source inside Claude and ChatGPT.
• Finyuus: code-first language for durable, governed AI workflows.

Vibe-coding

• Maple-Preview: ternary 20B MoE model running at 120 tok/s on-device on iPhone; HN notes confident hallucination is the real blocker for small on-device models, they need reliable tool-calling since they can't memorize facts or flag their own uncertainty.