CEREBRO

Researched and written with 304,041 tokens (in 10 · out 29,534 · cache-read 65,545 · cache-create 208,952) across 5 Claude calls

Coding agents

headroomlabs-ai/headroom compresses tool outputs, logs, files and RAG chunks before they hit the LLM (about 20% fewer tokens for coding agents, 60-95% for JSON), shipped as a library, proxy and MCP server, so it is a drop-in context-cost cut.
v2.1.304 of Claude Code adds decider: "person" to a mod's tool.check answer, so only the permission dialog or a host callback can approve, plus a transcript line when an earlier plugin skips one of your mod's hooks.
Anthropic's internal report: 80% of Claude session time is agents waiting on humans says human approval latency, not model speed, is the bottleneck, which argues for better permission and autonomy design.
Sandboxed Claude agents wrote 159 of my 188 PRs is a field report on running mostly-autonomous agents in sandboxes at real PR volume.
Launch HN: Laminar (YC S24) trains a model with RL to find errors in agent traces; the founders argue cheap models only need to flag obviously flawed steps, since propagated dumb errors cause most agent failures.
I built a Claude Code plugin that hands your session to the cloud when you close your MacBook, and teleports it back when you open it keeps long sessions running across laptop sleep.
We built a free browser canal game with Claude Code, and the harness it wrote plays the game to the end before anything ships is an example of agent-written end-to-end verification as a ship gate.

RT @sentient_agency: OpenCode users' AI keys stolen says an attacker broke into a legacy system that OpenCode was shutting down on October 16 and took OpenAI, Anthropic and Google keys users had saved in its paid Console dashboard; users reportedly saw it on their bills first, a reminder not to park provider keys with a hosted agent vendor.

GitHub Copilot models deprecated on October 19, 2026 retires several models across Chat, inline edits, ask and agent modes and completions, so check any pinned model IDs.
Copilot model policy API for enterprise, orgs, and teams adds REST GET and PATCH endpoints to read and set which models are allowed per enterprise, team, app or org.
How to continue a GitHub Copilot app session in VS Code shows how to move an agent session into VS Code while keeping the conversation, code changes and working context.

Starkomand runs AI coding agents on an infinite canvas.
CrawlForge MCP gives an agent 31 web tools through one MCP server.
walkai-talkai lets you talk to your coding agent over a walkie-talkie interface.
Lancetta tells you when to push your coding agents and when to ease off.

Talk by Anthropic's Head of Coding Agents on vibe coding is a 30-minute Spanish-language post pointing to a talk on vibe coding.
I rebuilt a canvas app AI short film workflow inside one Claude Code session shows a creative multi-step pipeline run inside a single Claude Code session.