subagent-driven-development

The recommended execution engine: run a plan by dispatching a fresh implementer subagent per task, a task review after each, and a broad whole-branch review at the end — all in the current session, without pausing for human check-ins between tasks (pattern-fresh-context-subagents).

Core principle: fresh subagent per task + task review (spec compliance ∥ code quality) + broad final review = high quality, fast iteration. Subagents never inherit the controller’s context; the controller crafts exactly the brief each one needs, preserving its own context for coordination.

The two-stage task review is the pattern-adversarial-review gate: the reviewer returns both a spec-compliance verdict (nothing missing, nothing extra) and a code-quality verdict; Critical/ Important findings are fixed by a dispatched fix subagent and re-reviewed before the task is marked complete → artifact-atomic-commit per task. The final whole-branch review uses sp-requesting-code-review’s template, then work closes via sp-finishing-a-development-branch. Implementer subagents follow sp-test-driven-development.

Sharp operational discipline distinguishes it: explicit model selection per role (cheapest tier for transcription-style tasks, most capable for the final review — never inherit the session model); file handoffs (task briefs, reports, and review packages move as files, never pasted into context); a durable progress ledger (.superpowers/sdd/progress.md) so a compaction never re-dispatches a completed task; and a strict rule against pre-judging reviewer findings (“do not flag X” is forbidden). It clusters with the other fresh-context executors — gsd-execute-phase (wave-based executors) and bmad-dev-story (one fully-contexted story per fresh context).

See Also