Sub-agent context isolation: return 1–2K-token summaries to the lead agent, not full transcripts

Arize AI·high signal

Give each subagent its own isolated context window and have it return only a condensed 1,000–2,000 token summary to the orchestrator instead of its raw working context. This keeps the lead agent's window clean across long horizons and is the architecture behind Claude Code's fork/subagent pattern. Use it when a subtask reads many files or needs specialized focus.