This page documents the subagent-driven development (SDD) workflow, which executes implementation plans by dispatching fresh subagents per task with mandatory two-stage review. The controller agent orchestrates implementer subagents and reviewer subagents, managing the task queue and ensuring quality gates pass before marking tasks complete.
The SDD workflow is the final execution phase of the Superpowers pipeline, following brainstorming and plan writing. It ensures high-quality output through isolated contexts and specialized model selection.
Sources: skills/subagent-driven-development/SKILL.md1-12 skills/subagent-driven-development/SKILL.md40-60
Decision criteria:
| Criterion | Required for SDD |
|---|---|
| Implementation plan exists | Yes - from writing-plans skill |
| Tasks are independent | Yes - can be implemented in any order |
| Subagent support | Yes - harness must support Task tool (Claude Code, Codex) |
| Session continuity | Yes - work continues in current session |
SDD is the preferred execution method when subagents are available because it prevents context pollution and allows for specialized review loops. If subagents are not available, the system falls back to the executing-plans skill.
Sources: skills/subagent-driven-development/SKILL.md14-39
The SDD architecture separates the "Controller" (which maintains the global plan state) from "Subagents" (which perform isolated units of work).
Key architectural decisions:
.superpowers/sdd/, a self-ignoring directory managed by sdd-workspace skills/subagent-driven-development/scripts/sdd-workspace1-23task-reviewer-prompt.md dispatch to reduce overhead while maintaining two distinct verdicts docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md54-59Sources: skills/subagent-driven-development/SKILL.md10-12 skills/subagent-driven-development/scripts/sdd-workspace1-23 docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md54-59
Critical invariants:
sdd-workspace to minimize controller context bloat skills/subagent-driven-development/scripts/task-brief1-8Sources: skills/subagent-driven-development/SKILL.md45-83 skills/subagent-driven-development/SKILL.md85-97 skills/subagent-driven-development/task-reviewer-prompt.md42-47 skills/subagent-driven-development/scripts/task-brief1-8
The workflow utilizes different model tiers to balance cost, speed, and reasoning capability.
| Task Type | Model Tier | Complexity Signals |
|---|---|---|
| Mechanical | Cheap / Fast | Isolated functions, clear specs, 1-2 files. Plan contains complete code skills/subagent-driven-development/SKILL.md103 |
| Integration | Standard / Mid-tier | Multi-file coordination, pattern matching, debugging skills/subagent-driven-development/SKILL.md105 |
| Architecture | Most Capable | Design judgment, broad final review, complex concurrency skills/subagent-driven-development/SKILL.md107-109 |
Key Principle: "Turn count beats token price." Cheaper models often take 2-3x more turns on multi-step work, potentially costing more than mid-tier models in wall-clock time and context skills/subagent-driven-development/SKILL.md119-122
Sources: skills/subagent-driven-development/SKILL.md99-125
The SDD workflow relies on several shell scripts to manage data flow between subagents.
| Script | Purpose | Output Path |
|---|---|---|
sdd-workspace | Ensures the .superpowers/sdd/ directory exists with a * gitignore skills/subagent-driven-development/scripts/sdd-workspace1-23 | .superpowers/sdd/ |
task-brief | Extracts the specific text for Task N from the implementation plan skills/subagent-driven-development/scripts/task-brief1-25 | .superpowers/sdd/task-N-brief.md |
review-package | Generates a combined file containing commit logs, stats, and the full diff skills/subagent-driven-development/scripts/review-package1-41 | .superpowers/sdd/review-BASE..HEAD.diff |
Sources: skills/subagent-driven-development/scripts/sdd-workspace1-23 skills/subagent-driven-development/scripts/task-brief1-25 skills/subagent-driven-development/scripts/review-package1-41
The task reviewer (task-reviewer-prompt.md) provides a unified gate for each task.
Reviewers categorize findings by severity:
Reviewers are instructed: "Do Not Trust the Report." They must verify implementer claims against the actual diff and treat implementer design rationales (e.g., "left it per YAGNI") as claims to be judged, not verdicts skills/subagent-driven-development/task-reviewer-prompt.md55-62
Sources: skills/subagent-driven-development/task-reviewer-prompt.md1-166 docs/superpowers/specs/2026-06-09-sdd-task-scoped-review-dispatch-design.md60-68
Implementer subagents report one of four statuses. The controller must handle each appropriately:
| Status | Meaning | Controller Action |
|---|---|---|
| DONE | Task complete | Proceed to task review skills/subagent-driven-development/SKILL.md129 |
| DONE_WITH_CONCERNS | Work complete but implementer has doubts | Address concerns; proceed to review if they are minor skills/subagent-driven-development/SKILL.md131 |
| NEEDS_CONTEXT | Missing information | Provide missing context and re-dispatch skills/subagent-driven-development/SKILL.md133 |
| BLOCKED | Cannot complete task | Assess: upgrade model, break down task, or escalate to human if plan is broken skills/subagent-driven-development/SKILL.md135-139 |
Sources: skills/subagent-driven-development/SKILL.md127-141 skills/subagent-driven-development/implementer-prompt.md127-138