# /subagent-driven-development: Subagent-driven plan execution with per-task review

Source: https://trendhub.net/skills/obra-superpowers/subagent-driven-development
Skill folder: https://github.com/obra/superpowers/tree/main/skills/subagent-driven-development
Pack: Superpowers (https://trendhub.net/skills/obra-superpowers)
Works with: Claude Code, Claude Code (Superpowers marketplace), Codex CLI, Cursor, Gemini CLI, GitHub Copilot CLI, Devin CLI, Qwen Code, Pi, Antigravity

Subagent-driven development executes an implementation plan by giving each task to a fresh subagent with only the context it needs, then reviewing the result for spec compliance and code quality. Progress is tracked in a ledger file so the work survives context compaction, and a whole-branch review runs at the end.

## When to use it

- You have an approved plan whose tasks are mostly independent
- Your agent supports subagents and you want thorough review after every task
- A long session needs to keep going for hours without drifting from the plan

## How it works

1. **Set up the workspace**: It confirms an isolated worktree, creates a per-plan ledger under .superpowers/sdd/ and scans the plan for conflicts.
2. **Dispatch an implementer**: Each task goes to a new subagent that implements, tests, commits and self-reviews.
3. **Review the task**: A reviewer subagent checks the change against the spec and for quality, with a fix loop when issues are found.
4. **Record and continue**: Completion and any judgment calls are written to the ledger, then the next task starts without pausing.
5. **Final review**: After all tasks, a broad review of the whole branch runs before handing off to finishing-a-development-branch.

## What it steers the agent away from

- Re-running already finished tasks after the agent loses context
- Stopping to ask for permission between every task
- Starting implementation directly on the main branch without consent

## What it can touch

Runs shell commands and a bundled workspace script, creates and edits code files, runs tests and makes git commits in the working tree, and writes ledger and review files in a git-ignored .superpowers/sdd/ directory. It stops to ask before destructive, security-sensitive or shared actions such as merges, pushes or publishing.

## Pairs well with

- `/writing-plans`
- `/executing-plans`
- `/using-git-worktrees`
- `/finishing-a-development-branch`

## Questions

### How is it different from executing-plans?

Executing-plans does every task inline in one context with a single review at the end, which costs less; subagent-driven development uses a fresh context per task and reviews each one, which is more thorough.

### Which models does it use?

It recommends the least capable model that can handle each role: cheaper models for mechanical tasks and the most capable one for design work and the final review.
