Subagent development

Subagent-Driven Development

Fresh subagent per task + review between tasks = high quality, fast iteration.

When to Use

• Executing implementation plans in current session
• Tasks are largely independent
• Want continuous progress with quality gates
• Need isolation between task contexts

When NOT to Use

• Plan needs initial review/revision
• Tasks have tight sequential dependencies
• Simple changes that don't need isolation

The Process

1. Load Plan & Create Tasks

1. Read the plan document
2. Create TodoWrite with all tasks
3. Verify plan is sound before starting

2. For Each Task

Dispatch Implementation Subagent:

Task([appropriate-agent], prompt="
Implement Task N from the plan:

**Task:** [Task description]

**Requirements:**
- Follow TDD (write failing test first)
- Implement minimal code to pass
- Verify all tests pass
- Commit with descriptive message

**Output:**
- What was implemented
- Test results
- Any issues encountered
")

3. Code Review Checkpoint

After EACH task, dispatch reviewer:

Task(code-reviewer, prompt="
Review implementation of Task N:

**What was implemented:** [From subagent output]
**Commits:** [Range]
**Requirements:** [From plan]

Evaluate:
- Implementation vs requirements match
- Code quality and patterns
- Test coverage
- Critical/Important/Minor issues
")

4. Address Feedback

Severity	Action
Critical	Fix immediately, re-review
High	Fix before next task
Medium	Fix if quick, else document
Low	Document for later

5. Mark & Continue

1. Update TodoWrite to mark task complete
2. Proceed to next task
3. Repeat cycle

6. Final Review

After all tasks complete:

Task(code-reviewer, prompt="
Final review of complete implementation:

**Plan:** [Reference]
**All commits:** [Full range]

Verify:
- Complete plan compliance
- Architectural soundness
- No regressions
- Ready for merge
")

7. Finish

Transition to merge/PR workflow.

Critical Rules

Never:

• Skip code reviews between tasks
• Proceed with unresolved Critical issues
• Run multiple implementation subagents in parallel
• Implement without plan task reference

Always:

• Fresh subagent for each task (clean context)
• Review after each task (quality gate)
• Fix Critical/High before continuing
• Document deferred issues

Quality Gate Template

## Task N Review

**Status:** [Pass/Fail]

### Critical Issues

[None / List]

### High Priority

[None / List]

### Medium Priority

[None / List]

### Minor/Suggestions

[None / List]

### Verdict

- [ ] Proceed to next task
- [ ] Fix issues first
- [ ] Needs discussion

Benefits

• Clean context for each task (no accumulated confusion)
• Built-in quality gates catch issues early
• Reviewable progress at each step
• Easy to pause/resume at task boundaries