Subagent-Driven Development

Dispatch fresh subagent per task for fast, isolated execution with quality gates.

When to Use

• 3+ independent issues to investigate
• Implementation plan with independent tasks
• Need parallel execution without context pollution
• Multiple test failures across different files

Sequential Execution

For tightly coupled tasks that must run in order:

For each task:
  1. Dispatch implementation subagent
  2. Review output
  3. Fix issues if found
  4. Mark complete, move to next

Parallel Execution

For independent tasks (different files, subsystems):

1. Identify independent domains
2. Dispatch one agent per domain
3. Review all results
4. Verify no conflicts
5. Run full test suite

Subagent Prompt Template

You are implementing [TASK] from [PLAN].

Your job:
1. Read the requirements carefully
2. Implement exactly what's specified
3. Write tests (TDD if specified)
4. Verify implementation works
5. Commit your work
6. Report back

Constraints:
- Don't change code outside your scope
- Follow existing patterns
- Use project's test framework

Report: What you implemented, tests run, files changed, any issues.

Review After Each Agent

Review the changes from the last subagent:
- Check for correctness
- Verify tests pass
- Look for scope creep
- Check code quality
- Report issues by severity: Critical/Important/Minor

Rules

• Fresh subagent per task — no context pollution
• Review between tasks — catch issues early
• Don't parallelize coupled tasks — conflicts will occur
• Stop on blockers — don't guess, ask
• Verify full test suite — after all agents complete

When NOT to Use

• Tasks are tightly coupled
• Need full system context to understand
• Exploratory debugging (don't know what's broken)
• Agents would edit same files