Claude Code sessions can now talk to each other
One session passes another a short summary rather than the full chat history and files, and the second session picks it up mid-task. It works both ways: you can ask a neighboring session a question and get the answer back in your active one. Per Anthropic’s description, Claude also sends messages on its own, for instance when an edit it just made affects what another session is doing.
The messages travel through a mailbox mechanism inside agent teams, the experimental feature for running several sessions together. It’s off by default. You turn it on with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json or in your environment, and Anthropic’s docs walk through the setup. macOS and Linux only, after updating Claude Code.
Up next
By the end of a long session, strict solutions drop to 0.5%
SlopCodeBench had 11 models iterate on their own code across 93 checkpoints: no agent ever finished a task, the best checkpoint score was 17.2%, and by the final checkpoint strict solutions fell to 0.5% while cost rose 2.9x. Structural degradation showed up in 80% of runs.
Related stories
- agent-mesh lets coding agents talk to each other over MCP
- Claude Code has four kinds of agent loops: /goal, /loop, /schedule
- Letting Claude filter its own tool output took Opus 4.6 from 45.3% to 61.6% on BrowseComp
- A hardcoded instruction in Claude Code 2.1.219 and 2.1.220 blocks Opus 5 from spawning sub-agents
- Analysis of the agent control model in Claude Code
- Claude integrates with Lovable
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.