Subagent architecture coordinates specialized agents across a larger workflow instead of extending one conversation until context decays. Each task receives a fresh context window, while durable project state moves into files such as CONTEXT.md and a .planning/ directory. section 3.5.3
A disciplined implementation follows four stages:
- Discuss: clarify product decisions, expose assumptions, and persist the answers in CONTEXT.md. section 3.5.3
- Plan: research the codebase and divide implementation into small tasks that a fresh agent can execute with zero assumed codebase context. section 3.5.3
- Execute: assign each task to a specialized subagent with a clean context window rather than compacting one growing conversation. section 3.5.3
- Verify: preserve plans, research notes, validation contracts, and task-level Git commits so failures can be traced against project state. section 3.5.3
