subagent-driven-development

Executing an implementation plan with parallel subagents involves dispatching a fresh subagent per task using dependency-aware waves. The system computes ready-sets and validates file-overlap to ensure safe parallel execution, collecting self-review concerns for later audit.

Dependency-aware wave-based execution groups independent development tasks to run in parallel, computing ready-sets and applying cycle checks per wave. Tasks with dependencies wait, ensuring safe parallel execution while accelerating iteration across the implementation plan.

Subagents handle self-review by evaluating their own task implementation before returning a structured status summary. They collect any concerns encountered during execution, logging them into an audit trail for later review and deterministic development traceability.

File-overlap validation prevents conflicts during parallel task execution by checking if independent subagents target the same files. It computes ready-sets per wave and blocks overlapping tasks, ensuring safe parallel execution without write conflicts.

Structured status summaries report task completion states: DONE indicates success, DONE_WITH_CONCERNS flags issues for audit, NEEDS_CONTEXT requests missing information, and BLOCKED shows unresolvable dependencies, enabling deterministic development tracking.

You should not use subagent-driven parallel execution when tasks have heavy file-overlap or circular dependencies that prevent wave-based separation. If cycle checks detect unresolvable blockers, sequential execution is required to ensure safe development.