agent-workflow

Agent Workflow

When to Use

• Multi-step tasks that benefit from parallel execution
• Complex research or analysis
• Codebase-wide changes
• Tasks requiring specialized subagents

Quick Reference

Task	Approach
Parallel research	delegate_task with multiple goals
Code implementation	delegate_task with TDD context
Review/QA	delegate_task with review checklist
Long-running	terminal(background=true) + process

The Plan-Execute-Verify Loop

1. Plan

• Break task into 3-7 subtasks
• Identify dependencies between subtasks
• Determine which can run in parallel
• Assign tools and constraints per subtask

2. Execute

# Parallel execution for independent tasks
delegate_task(tasks=[
    {"goal": "Research X", "context": "..."},
    {"goal": "Research Y", "context": "..."},
    {"goal": "Research Z", "context": "..."},
])

# Sequential for dependent tasks
result_1 = delegate_task(goal="Step 1")
result_2 = delegate_task(goal="Step 2", context=result_1)

3. Verify

• Check each subtask output against success criteria
• Integration test the combined result
• Document what worked and what didn't
• Update memory/lessons learned

Subagent Patterns

Investigator → Builder → Reviewer

# 1. Investigate
investigation = delegate_task(
    goal="Investigate root cause of auth token expiry",
    context="User reports 401 errors after 30 minutes. Stack: Next.js, Supabase.",
)

# 2. Build
build = delegate_task(
    goal="Fix the auth token refresh issue",
    context="Root cause: {investigation}. Follow TDD. Use Supabase auth.",
)

# 3. Review
review = delegate_task(
    goal="Review the fix for correctness and security",
    context="Fix: {build}. Check: token refresh, CSRF, session fixation.",
)

Parallel Scrub

# When order doesn't matter, fan out
delegate_task(tasks=[
    {"goal": "Update README", "context": "..."},
    {"goal": "Add tests", "context": "..."},
    {"goal": "Fix lint errors", "context": "..."},
    {"goal": "Update types", "context": "..."},
])

Pitfalls

• Over-parallelizing → Some tasks need sequential context
• Vague goals → Subagents need specific, checkable objectives
• No verification → Always check subagent output
• Ignoring failures → One failed subtask can invalidate the whole
• Too many levels → Max 2 levels of delegation

Verification

• [ ] All subtasks completed successfully
• [ ] Outputs integrated correctly
• [ ] Success criteria met for each subtask
• [ ] No regressions from parallel changes
• [ ] Lessons learned documented
• House rules: docs/house/HOUSE-MOTION-RULE.md and docs/standards/AGENT-STANDARD.md. Read before any design, site, dashboard or header job. Pointers only; do not restate rules here.
• Current beats existing: fetch and pull main first; origin main wins over clone, cache, journal and definition file; a clone that cannot fast-forward stops and reports (ops/handoffs/FABLE-HANDOFFS-2026-11-06.md section 0).