Multi Agent Autonomous Workflow

Long-Running Autonomous Workflow

Run for hours until the goal is complete. The Stop hook blocks exit until done.

Core Execution Model

┌─────────────────────────────────────────────────────────────────┐
│                    LONG-RUNNING EXECUTION                        │
│                                                                  │
│  1. ANALYZE: Break goal into stories                            │
│  2. LOOP: For each story, iterate until verified                │
│  3. COMPLETE: All stories done → create PR → WORKFLOW_COMPLETE  │
│                                                                  │
│  Stop hook blocks exit until WORKFLOW_COMPLETE marker output    │
└─────────────────────────────────────────────────────────────────┘

Guiding Principles

1. Persist state - Survive any interruption
2. Verify before progressing - Build and tests must pass
3. Learn from failures - Use failures as context for retry
4. Escalate when stuck - Don't loop forever, ask for help
5. Think before acting - Reason about risks, don't just execute

Everything else is agent reasoning. You already know about flaky tests, merge conflicts, external dependencies, performance issues. Use that knowledge.

Workflow Steps

Phase 0: Recovery Check

python .claude/core/state.py recover

If resuming, continue from where you left off.

Phase 1: Analysis

1. Explore codebase first - Before asking questions, understand:

   • Project structure and architecture
   • Existing patterns and conventions
   • Related code that might be affected
   • What already exists vs what needs building

2. Ask informed questions (use AskUserQuestion tool)

   Based on codebase analysis, ask specific questions like:

   • "I see you have X pattern. Should the new feature follow this?"
   • "This will affect modules A, B, C. Any concerns?"
   • "I found existing Y. Should I extend it or build new?"

   Not generic questions like "What framework?" - you already know from the code.

3. Initialize workflow:

   python .claude/core/state.py init "Goal description"
   

4. Break goal into stories:

   python .claude/core/state.py add-story "Story title" --size M
   

Phase 2: Story Loop

For each story:

1. Start story:

   python .claude/core/state.py update S1 in_progress
   

2. Implement with TDD (write test → make it pass → refactor)

3. Verify (MANDATORY):

   # Get platform commands
   BUILD_CMD=$(python .claude/core/platform.py get-command build)
   TEST_CMD=$(python .claude/core/platform.py get-command test)
   
   # Run them
   eval "$BUILD_CMD" && eval "$TEST_CMD"
   

4. If verification fails: Don't exit. Use the failure as context. Try again.

5. If verification passes:

   python .claude/core/state.py verify S1 buildPasses --passed
   python .claude/core/state.py verify S1 testsPasses --passed
   python .claude/core/state.py update S1 completed
   

6. Repeat until no incomplete stories remain.

Phase 3: Completion

1. Final verification - Build + test must pass
2. Create PR with summary of changes
3. Complete workflow:

   python .claude/core/state.py complete
   

4. Output completion marker:

   ## WORKFLOW_COMPLETE
   
   PR created. All stories verified.
   

Handling Problems

Build/Test Fails

Don't exit. Analyze the error. Fix it. Try again.

Stuck After 3+ Attempts

python .claude/core/state.py blocker "Description of what's blocking"

Escalate to user or try a different approach.

Need Human Action

python .claude/core/state.py await-user "What needs to be fixed" --check "verification command"

Stop hook will allow exit. User fixes issue, runs user-fix-complete, restarts.

External Dependencies

Think about what's needed. Mock it for tests. If credentials required, use await-user.

State Commands

# Lifecycle
python .claude/core/state.py init "Goal"
python .claude/core/state.py status
python .claude/core/state.py recover
python .claude/core/state.py complete

# Stories
python .claude/core/state.py add-story "Title" --size M
python .claude/core/state.py update S1 in_progress
python .claude/core/state.py update S1 completed
python .claude/core/state.py verify S1 buildPasses --passed
python .claude/core/state.py next

# Problems
python .claude/core/state.py blocker "Description"
python .claude/core/state.py await-user "Description"
python .claude/core/state.py user-fix-complete

# Git
python .claude/core/state.py checkpoint "Before risky change"
python .claude/core/state.py rollback

Platform Commands

python .claude/core/platform.py detect      # Auto-detect platform
python .claude/core/platform.py get-command build
python .claude/core/platform.py get-command test

Stop Hook Behavior

The Stop hook (stop.py) implements persistent execution:

• On exit attempt: Checks for WORKFLOW_COMPLETE marker
• If not found: Blocks exit, provides context to continue
• If found: Allows normal exit
• Safety limit: Max 100 iterations prevents infinite loops
• User intervention: If awaiting_user, allows exit with instructions

Key Points

• Never exit without WORKFLOW_COMPLETE (stop hook enforces)
• Never skip verification (build + test before completing story)
• Iterate on failures (failures are context, not blockers)
• Use your reasoning (you know more than any checklist)