LLM mechanics: caching, tokens, routing

• KV-cache hit rate is the #1 production agent metric — Manus AI's verdict, with LangChain amplifying: if your agent isn't measuring this, you're flying blind on cost and latency.
• headroomlabs-ai/headroom — library/proxy/MCP server that compresses tool outputs, logs, and RAG chunks 60-95% before they hit the LLM; plug-in, not a rewrite.
• workweave/router — localhost proxy that picks Anthropic/OpenAI/Gemini model per-request via on-box embedder; community flag: routing breaks prompt caching so net cost may rise.

Coding agents & IDEs

• Letta Mods: agents rewriting their own harness — agents can now modify the execution harness they run in (inject tools, customize context, statusline); inspired by Pi's extension system.
• MAI-Code-1-Flash GA for Copilot Business/Enterprise — Microsoft's in-house coding model now generally available; another non-OpenAI model in the Copilot fleet.
• GitHub Desktop 3.6 — Git worktree support lands in Desktop alongside Copilot-powered commit authoring and merge conflict resolution.
• T3 npx t3@nightly serve over Tailscale — remote coding server now works cleanly over Tailscale; Theo does 90% of coding from browser.
• xAI/SuperGrok subscriptions usable inside T3 Code — Grok models now accessible via T3 Code with existing X/SuperGrok subscription.
• BigPizzaV3/CodexPlusPlus — external launcher for Codex App that injects enhancements via CDP without touching the original install; Windows/macOS.
• Claude Code v2.1.195 — hook matchers now exact-match hyphenated identifiers (no more code-reviewer accidentally matching code-reviewer-extra); CLAUDE_CODE_DISABLE_MOUSE_CLICKS env var added.

Models & research

• Claude Mythos 5 cleared for 100+ US institutions — US government lifted its block; Mythos (fewer guardrails) approved while Fable (more guardrails) was denied — contradicts the stated safety rationale.
• MIT: code volume +300%, shipped output +30% — study of 100k+ GitHub devs finds AI agents inflate commits 180% but releases only rise 30%; the gap is the quality/review bottleneck, not generation.
• Frontier open-source LLM gap analysis — quantifies the performance gap between open-weights and closed models; predicts a frontier open-weights release by Dec 2026; community notes sustainability risk given DeepSeek-style corporate dependency.

Agentic frameworks & SaaS

• BuilderIO/agent-native — open-source framework for agents acting inside real apps: shared actions, SQL state, identity, tools, skills, jobs, observability, UI surfaces — BYOD (database/hosting/model).
• Cursor delegated tasks from Notion — @Cursor on a Notion spec opens a PR via the Cursor SDK cloud agent; same models/harness as local Cursor.
• Claude-to-Claude-Code context tool — community-built MCP/tool letting a Claude chat session query Claude Code for live codebase context; closes the context gap between chat and agentic IDE.
• Agent Arena — first public benchmarking arena specifically for AI agents head-to-head.
• Atlas (Nanonets) — company-context layer so every AI tool you use knows how your org works; cross-tool memory/knowledge graph play.
• LockIn MCP — MCP server that lets your agent block OS-level distractions on demand; shallow but practical focus tool.

CLI/TUI

• BrowserBash — CLI that converts plain-English into real browser test scripts; natural-language Playwright wrapper aimed at non-test engineers.