Sub-agents

The agent tool spawns a fresh agent loop with its own session id, tool set, and iteration budget, fully isolated from the parent’s history. Read-only roles run in explore mode and write-capable roles in build mode; only the final result comes back to the parent turn.

On this page (coming soon)

• The four roles: explorer, reviewer, tester, summarizer
• Why isolation keeps the parent’s context small
• Iteration budgets and capability profiles
• Watching progress via subagentStarted / subagentCompleted
• When delegation helps and when it just costs tokens