Multi-Agent Workflow
Overview
This skill enables sophisticated parallel development workflows using multiple Claude instances working concurrently in separate git worktrees. The workflow consists of a manager session (current session), worker agents performing parallel implementations, and observer subagents monitoring progress and collecting results.
Architecture
Manager Session (Current Session)
- Orchestrates the entire workflow
- Distributes tasks to multiple worker agents
- Evaluates implementation results
- Selects and merges the best implementation
Worker Agents (Git Worktrees)
- Execute in isolated git worktrees
- Implement the same task using different approaches
- Work independently and in parallel
- Each worker has its own branch
Observer Subagent
- Monitors worker progress and status
- Collects implementation results
- Performs initial quality assessment
- Reports consolidated findings to manager
When to Use This Skill
Use this skill when:
- Multiple implementation approaches should be explored simultaneously
- Complex tasks benefit from parallel A/B testing of different solutions
- You need to compare different algorithms or architectural patterns
- Time-sensitive development requires faster iteration through parallelization
- Exploring trade-offs between different technical approaches
Workflow Steps
1. Task Distribution
The manager session:
- Receives implementation task from user
- Breaks down into parallelizable work units
- Sets up git worktrees for each worker
- Defines evaluation criteria
2. Parallel Execution
Worker agents:
- Receive specific implementation approach/variant
- Work in isolated git worktrees
- Implement, test, and commit independently
- Each uses different strategy or optimization
3. Monitoring & Collection
Observer subagent:
- Monitors all worker sessions
- Tracks completion status
- Collects implementation artifacts
- Runs preliminary quality checks (tests, lint, etc.)
4. Evaluation & Selection
Manager session:
- Reviews all implementations
- Compares based on criteria:
- Code quality and maintainability
- Performance metrics
- Test coverage
- Complexity and readability
- Edge case handling
- Selects best implementation
- Merges into manager's target branch
Usage Instructions
Step 1: Initialize Multi-Agent Workflow
You (Manager): I need to implement [feature/fix] using multi-agent workflow.
Let's explore [N] different approaches:
1. Approach A: [description]
2. Approach B: [description]
3. Approach N: [description]
Evaluation criteria:
- [criterion 1]
- [criterion 2]
- [criterion N]
Step 2: Set Up Worktrees
The manager uses the setup script to create git worktrees:
./.claude/skills/multi-agent-workflow/scripts/setup-worktrees.sh \
--base-branch "claude/multi-agent-workflow-$(date +%s)" \
--num-workers 3 \
--task-id "feature-implementation"
This creates:
worktree-1/with branchworker-1/feature-implementationworktree-2/with branchworker-2/feature-implementationworktree-3/with branchworker-3/feature-implementation
Step 3: Launch Worker Agents
The manager launches worker Claude instances in each worktree using Task tool:
Send a single message with multiple Task tool calls:
- Task 1: Worker agent 1 - Implement using approach A
- Task 2: Worker agent 2 - Implement using approach B
- Task 3: Worker agent 3 - Implement using approach C
- Task 4: Observer agent - Monitor all workers and collect results
Each worker receives:
- Task description
- Specific implementation approach
- Evaluation criteria
- Worktree path to work in
Step 4: Observer Monitoring
The observer subagent:
- Periodically checks worker progress
- Collects completed implementations
- Runs automated quality checks
- Compiles comparison report
Step 5: Evaluation and Selection
Manager receives observer report containing:
- Implementation summaries
- Test results
- Performance metrics
- Code quality analysis
Manager evaluates and selects best implementation:
./.claude/skills/multi-agent-workflow/scripts/evaluate-results.sh \
--worktrees "worktree-1 worktree-2 worktree-3" \
--criteria "performance,maintainability,test-coverage" \
--output "evaluation-report.md"
Step 6: Accept Best Implementation
Manager merges selected implementation:
git checkout claude/multi-agent-workflow-[timestamp]
git merge --no-ff worker-2/feature-implementation
git worktree remove worktree-1 worktree-2 worktree-3
Example Scenarios
Scenario 1: Algorithm Optimization
Task: Implement data processing pipeline Workers:
- Worker 1: Stream processing approach
- Worker 2: Batch processing approach
- Worker 3: Hybrid approach Criteria: Performance, memory usage, code simplicity
Scenario 2: UI Component Design
Task: Create user authentication form Workers:
- Worker 1: Material UI components
- Worker 2: Custom styled components
- Worker 3: Headless UI with Tailwind Criteria: Accessibility, bundle size, maintainability
Scenario 3: API Implementation
Task: Build REST API endpoint Workers:
- Worker 1: Express with middleware pattern
- Worker 2: Fastify with schema validation
- Worker 3: Next.js API routes Criteria: Performance, type safety, error handling
Scenario 4: Bug Fix Exploration
Task: Fix complex race condition Workers:
- Worker 1: Mutex locks approach
- Worker 2: Event queue approach
- Worker 3: Refactor to eliminate shared state Criteria: Correctness, simplicity, performance impact
Best Practices
Task Design
- Clear Constraints: Define exact scope and boundaries for each worker
- Distinct Approaches: Ensure each worker uses meaningfully different strategy
- Measurable Criteria: Set objective, quantifiable evaluation metrics
- Realistic Scope: Size tasks appropriately for parallel completion (30min-2hrs)
Worker Management
- Isolation: Ensure workers don't interfere with each other
- Clear Prompts: Provide specific implementation guidance to each worker
- Resource Limits: Consider computational resources when setting worker count
- Status Updates: Request periodic progress updates from workers
Observer Configuration
- Monitoring Frequency: Balance between responsiveness and overhead
- Quality Gates: Define automated checks (tests pass, lint clean, etc.)
- Reporting Format: Structure observer reports for easy comparison
- Early Termination: Allow observer to flag critical failures early
Evaluation Process
- Objective Metrics: Prioritize quantifiable measurements
- Code Review: Manually review final candidates
- Testing: Run comprehensive test suites on all implementations
- Documentation: Require workers to document their approach
- Hybrid Solutions: Consider combining best aspects of multiple implementations
Implementation Templates
Worker Prompt Template
Located at: .claude/skills/multi-agent-workflow/templates/worker-prompt.md
Observer Prompt Template
Located at: .claude/skills/multi-agent-workflow/templates/observer-prompt.md
Scripts Reference
setup-worktrees.sh
Creates git worktrees for parallel development.
Usage:
./scripts/setup-worktrees.sh --base-branch <name> --num-workers <n> --task-id <id>
launch-worker.sh
Launches Claude worker instance in specified worktree.
Usage:
./scripts/launch-worker.sh --worktree <path> --task <description> --approach <strategy>
evaluate-results.sh
Evaluates and compares worker implementations.
Usage:
./scripts/evaluate-results.sh --worktrees <paths> --criteria <metrics> --output <file>
Troubleshooting
Worker Conflicts
If workers modify the same files unexpectedly:
- Review task scope definitions
- Ensure worktrees are properly isolated
- Check that workers are using correct branches
Observer Failures
If observer can't collect results:
- Verify worker completion status
- Check file paths and permissions
- Review observer monitoring configuration
Evaluation Ambiguity
If no clear best implementation emerges:
- Refine evaluation criteria
- Consider hybrid approach combining strengths
- Run additional targeted comparisons
- Consult with user on priorities
Tips
- Start with 2-3 workers for first multi-agent workflow
- Document evaluation criteria before launching workers
- Use observer to catch issues early rather than waiting for completion
- Consider running test suites as part of automated evaluation
- Keep implementation approaches distinct enough to yield meaningful comparison
- Archive all worker implementations for future reference, not just selected one
Integration with Other Skills
- parallel-explore: Use before multi-agent-workflow to identify implementation approaches
- code-review: Apply to all worker implementations before final selection
- testing: Ensure each worker includes comprehensive test coverage
Scan to join WeChat group