How about using CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS in GSD? Thoughts on that?

[markuswahl-2]

Eg during the research part of the new porject command the numer is creeping up to 80% pretty quick....

[trek-e]

GSD already solves the context problem you're describing — each subagent gets a fresh 200K context window. The orchestrator stays thin and spawns focused agents, so context never fills up.

v1.31.0 also adds claim provenance tagging in the researcher, which tracks where information comes from so context is used more efficiently.

The Claude Code AGENT_TEAMS API was evaluated (PR #1539) but rejected — it introduces shared mutable state between agents, which is exactly what causes context rot.

---

Generated by Claude Code