| name | subagent-driven-development |
| description | Execute a plan through focused background agents. Use for multi-step implementation work where tasks can be isolated and reviewed one by one. |
Subagent-Driven Development
Use fresh workers to keep each task small, isolated, and reviewable.
Good Fit
- You already have a plan
- Tasks are mostly independent
- You want parallel execution without context sprawl
Core Rules
- Do not start implementation without a task list.
- Give each worker the exact task text; do not make workers rediscover the plan.
- Do not run two workers on the same files at the same time.
- Review every completed task before declaring it done.
Workflow
1. Prepare the task list
Break the work into 2-5 minute units with:
- objective
- exact files to touch
- verification command
- completion criteria
2. Dispatch one worker per independent task
Use start_background_agent for tasks that can proceed independently.
Each worker prompt should include:
- the task objective
- exact files
- repo conventions that matter
- required tests or commands
- instructions to stop and report if scope expands
3. Monitor and unblock
- Use
topology to see which agents are active.
- Use
send_agent_message(..., steering=true) if a worker is going down the wrong path.
- Use
agent_ask when you need a direct answer from a worker.
4. Review before merging the task
After each worker completes:
- Check spec compliance against the plan
- Run a code review using the
code-review skill
- Fix issues before taking the next dependent task
Anti-Patterns
- Parallelizing overlapping file edits
- Letting a worker "figure out the task" from broad context
- Skipping review because the task looked small
- Moving to the next task with open correctness concerns