Coding agents

• Claude Code v2.1.212: /fork now clones your conversation into a new background session (its own row in claude agents) while you keep working; the old in-session subagent spawn is renamed /subtask — release notes
• Claude Code v2.1.211: adds --forward-subagent-text / CLAUDE_CODE_FORWARD_SUBAGENT_TEXT to surface subagent text and thinking in stream-json, plus hardens permission-preview rendering against bidirectional-override and look-alike unicode spoofing — release notes
• Claude Code v2.1.210: live elapsed-time counter on collapsed tool-call summaries so long calls don't look stuck, plus a startup warning steering Write/NotebookEdit/Glob permission rules toward Edit/Read — release notes
• Cautionary pair on Reddit: one user burned a full 5-hour Claude Code limit in 6m32s (10.26M tokens, zero lines shipped) — thread; another warns subagents can silently blow past cost/scope expectations if left unsupervised — thread

LLM mechanics & models

• GLM-5.2 marks a real capability step-change for open-weight agentic models, per interconnects.ai's ongoing threshold tracking — analysis
• Claude Fable 5 release note: Anthropic now routes silent model manipulation of AI-safety-research queries through a classifier, matching how other safety domains are handled — post

MCP / protocol

• MCP 2.0 (spec 2026-07-28) ships "Stateless MCP," the most significant spec change since launch, and already spawned two new client projects — writeup
• llm-mcp-client 0.1a0: Simon Willison's client built against the new stateless MCP spec — release

Agentic SaaS & infra

• GitHub ships copilot-sdk (Python/TS/Go/.NET/Java/Rust): embed Copilot's production-tested agent runtime, planning, tool invocation, orchestration, directly in your own app instead of building it yourself — github
• Alibaba open-sources open-code-review: hybrid deterministic-pipeline + LLM-agent reviewer battle-tested internally at scale, line-level comments, built-in NPE/thread-safety/XSS/SQLi ruleset — github
• datasette-agent 0.4a0 adds await context.browser_task(), letting agent tools execute code directly in the user's browser — release
• qm: multiplayer agent harness giving each employee an isolated workspace plus shared Slack/web collaboration; HN commenters flag it as hastily-shipped and single-machine-only versus Hermes/Openclaw's distributed model — github
• SnapState: persistent state layer for AI agent workflows — hn
• TraceLLM: OpenTelemetry-style tracing for production AI applications — producthunt
• Instatic: open-source, self-hosted agentic visual CMS (Webflow/Framer alternative) that outputs clean static pages — github
• worldmonitor: AI-synthesized real-time intelligence dashboard, 500+ feeds across 15 categories, dual 3D/WebGL map engines with cross-stream correlation — github

CLI/TUI

• witr: traces why a process, port, container, or file is currently running — producthunt

Vibe-coding & repos

• last30days-skill: agent skill that researches a topic across Reddit/X/YouTube/HN/Polymarket and synthesizes an upvote-weighted, grounded summary — github
• QwenPaw: personal AI assistant with three-layer memory (live context, verbatim history, self-evolving knowledge base via ReMe), deployable locally or in the cloud — github
• OmniRoute: free MIT AI gateway aggregating 43 provider free tiers (516 models) behind one endpoint, quota-aware fallback plus Caveman-style prompt compression claiming 15-95% token savings — github
• Orca: ADE for running multiple coding agents (Claude Code, Codex, OpenCode, Pi) side by side, each in its own git worktree, with unified usage tracking — github
• Show HN revisit: Imba, a 7-year-old full-stack language compiling to JS with memoized-DOM; commenters note it's tightly coupled to its own framework — hn

Industry commentary

• Latent Space's AIE World's Fair 2026 recap: the field's center of gravity shifted from "building with agents" to "building systems around agents" — post
• AINews roundup: Codex is adding 1M users/day; Addy Osmani's keynote covers what to keep doing as inference cost keeps falling — post