By name: multi-agent-implementation description: Use when executing an implementation plan with multiple mostly independent tasks that can be delegated and reviewed in parallel.
Multi-Agent Implementation
Routing
Use only when implementation tasks are mostly independent and can be delegated in parallel. Use implement for one coherent task, test-driven-implementation when the main need is independent behavioural test authorship before implementation, strategy-audit when delegation itself is risky, and code-review or pre-ship for final quality checks.
Dispatch fresh agents per task. Two-stage review (spec compliance, then code quality) catches issues early.
This skill is about task concurrency and review between delegated tasks. It does not replace strict red-green implementation from independently authored tests.
When to Use
You have:
- An implementation plan with defined tasks
- Tasks that are mostly independent
- Want to stay in the same session
If tasks are tightly coupled, execute sequentially yourself instead.
Per-Task Workflow
1. Dispatch implementer agent
2. Implementer asks questions? → Answer with full context
3. Implementer implements, tests, commits, self-reviews
4. Dispatch spec reviewer agent
5. Spec passes? No → implementer fixes → re-review
6. Dispatch code quality reviewer agent
7. Quality passes? No → implementer fixes → re-review
8. Mark task complete
After all tasks: dispatch a final reviewer across the full implementation.
Use the least expensive model that can handle the role. If the runtime exposes named model tiers, map them to this intent:
| Complexity | Examples | Model |
|---|---|---|
| Mechanical | Isolated functions, clear specs, 1-2 files | small/fast |
| Integration | Multi-file changes, pattern matching, debugging | standard |
| Architecture | Design decisions, complex reviews | strongest available |
Handling Implementer Status
| Status | Action |
|---|---|
| DONE | Proceed to spec review |
| DONE_WITH_CONCERNS | Read concerns. If about correctness/scope, address first. If observations, note and proceed. |
| NEEDS_CONTEXT | Provide missing context, re-dispatch |
| BLOCKED | See troubleshooting below |
BLOCKED Troubleshooting
- Context problem → Provide context, re-dispatch
- Needs more reasoning → Re-dispatch with more capable model
- Task too large → Break into smaller pieces
- Plan wrong → Escalate to user
Never force retry without changing something.
Prompt Templates
Implementer
You are implementing Task N: [task name]
## Task Description
[FULL TEXT of task — paste it, don't make the agent read a file]
## Context
[Where this fits in the system, dependencies, architecture decisions]
## Before You Begin
If anything is unclear about requirements, approach, dependencies, or assumptions — ask before starting.
## Your Job
1. Implement exactly what the task specifies
2. Write tests (TDD: failing test first, then implementation)
3. Verify: build succeeds, tests pass
4. Commit your work
5. Self-review: completeness, quality, discipline, testing
6. Report back with status
## Report Format
- **Status:** DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT
- What you implemented
- What you tested and results
- Files changed
- Self-review findings
- Issues or concerns
Spec Reviewer
You are reviewing whether the implementation matches the specification.
## What Was Requested
[FULL TEXT of task requirements]
## What Implementer Claims
[From implementer's report]
## Critical: Do Not Trust the Report
Read the actual code. Compare to requirements line by line.
Check for:
- Missing requirements
- Extra/unneeded work
- Misinterpretations of the spec
## Report
- PASS: Everything matches spec after code inspection
- FAIL: [List what's missing/extra/wrong with file:line references]
Code Quality Reviewer
You are reviewing code quality for Task N.
## What Was Built
[From implementer's report]
## Review Criteria
- Each file has one clear responsibility
- Well-defined interfaces between units
- Testable independently
- Follows existing project patterns
- No unnecessary complexity
- Tests cover behaviour, not implementation details
## Report
- **Strengths:** [What's good]
- **Issues:** [Critical / Important / Minor, with file:line]
- **Assessment:** PASS | NEEDS_FIXES
Rules
- Never dispatch multiple implementers in parallel (merge conflicts)
- Never skip reviews (spec OR quality)
- Never proceed with unfixed issues
- Always provide full task text to agents (don't make them read files)
- If a reviewer finds issues, the same implementer fixes them, then re-review
- If an agent fails, dispatch a fix agent with specific instructions — don't fix manually unless the runtime has no agent support