Coding agents & ecosystem

• Claude Code v2.1.211 adds --forward-subagent-text to surface subagent text/thinking in stream-json output, and fixes permission previews so bidirectional-override, zero-width, and look-alike characters can no longer visually spoof tool-input previews in chat channels.
• Claude Code v2.1.210 adds a live elapsed-time counter to collapsed tool-call summaries and a startup warning steering Write/NotebookEdit/Glob permission rules toward Edit/Read.
• Codex is adding ~1M users/day, up 10x to 7M in six months, with AINews still fact-checking whether it's overtaken Claude Code.
• 12 ways a Claude architecture decision goes wrong: failure-mode checklist compiled while prepping for Anthropic's Professional cert.
• GitHub Copilot's July rollups: Visual Studio gets an SDK-based agent with .NET/Azure expertise baked in, VS Code v1.127-1.131 improves agent workflows, review, and chat navigation.

LLM mechanics & token economics

• Where Claude Code tokens actually go: breakdown from parsing 185 personal transcripts, useful ground truth for context/caching tuning.
• GLM-5.2 is the step change for open agents: open-weight model crosses a capability threshold on agentic tasks specifically, paired with a new paper on open RL recipes for terminal agents.
• Ontologies Are So Back: AI engineers reviving semantic-web ontologies to keep probabilistic agents inside deterministic boundaries.

Agent tooling & repos

• stablyai/orca: ADE running Codex/Claude Code/OpenCode/Pi side-by-side in separate worktrees, one dashboard for usage and rate limits.
• mattpocock/skills: Matt Pocock's personal .agents skill library, small/composable/model-agnostic, positioned against process-owning frameworks like GSD/BMAD/Spec-Kit.
• affaan-m/ECC: harness performance-optimization system (skills, instincts, memory, security) for Claude Code/Codex/Opencode/Cursor; the repo's own README warns hard about third-party re-upload supply-chain risk.
• Greplica: self-updating wiki generated for coding agents, keeps repo context current without hand-maintained docs.
• chrome-devtools-mcp: official Chrome DevTools MCP server, gives agents live browser control/inspection for debugging and perf work.
• alibaba/open-code-review: Alibaba's internal review tool open-sourced, deterministic pipeline + LLM agent hybrid with a fine-tuned ruleset (NPE, thread-safety, XSS, SQLi), OpenAI/Anthropic compatible.
• mvanhorn/last30days-skill: Claude Code skill that researches a topic across Reddit/X/YouTube/HN/Polymarket and synthesizes a grounded summary.
• huggingface/speech-to-speech: modular local voice-agent pipeline (VAD→STT→LLM→TTS) over an OpenAI Realtime-compatible WebSocket, LLM slot swappable to any OpenAI-compatible endpoint.
• Distillation censorship doesn't transfer: distilling DeepSeek into GPT-OSS doesn't carry over Chinese-model censorship, undercuts a common argument against Chinese open weights.

CLI/TUI

• llm 0.32rc2: Simon Willison's llm CLI fixes an rc1 dependency bug and switches the unset default model.
• llm-chat-completions-server 0.1a0: local server exposing llm's new content-addressable logs through an OpenAI Chat Completions-compatible API.
• tuicr: vim-keybinding TUI for code review, continuous diff scroll, line/file/range comments, exports to GitHub/GitLab/clipboard.

Agentic SaaS

• agentOS: WebAssembly-based sandbox claiming a 254x cost reduction over conventional agent execution environments.
• SnapState: persistent state layer for AI agent workflows, targets the session/context-continuity gap agents hit across restarts.
• LangWatch: Claude Code usage tracking: cost visibility tool showing what Claude Code sessions actually spend.