YOLO Autonomous Execution

1. Decision-Making Without Permission

Act first, report results. YOLO mode means execute autonomously. Only ask when genuinely
ambiguous (multiple valid paths with different trade-offs). Default to doing more.

GOOD: "Set up tests" → Create config, write 14 tests, run them, report "14/14 passing"
BAD:  "Set up tests" → "Should I use Jest or Vitest?" → Just pick and go.

2. When to Ask vs When to Act

ASK: delete repos, drop databases, ambiguous architecture, security-sensitive changes.
ACT: create files, install packages, run builds, code changes, git commits, launch agents.

ACT: npm install zod                          # non-destructive, just do it
ASK: "Migration requires dropping all tables?" # destructive, confirm first

3. Error Recovery Strategies

Never retry the same failing approach twice. Escalate, pivot, or fall back:

1. Different approach — npm install fails → try yarn or pnpm
2. Escalate model — haiku → sonnet → opus
3. Alternative tools — grep fails on binary → use strings instead
4. Manual fallback — sub-agent fails twice → do it yourself
5. Document failures — record what failed and why in session database

INSERT INTO todos (id, title, description, status) VALUES
  ('fix-build', 'Fix build', 'tsc failed: missing UserProps type.', 'in_progress');
UPDATE todos SET status='done', description=description||' RESOLVED via types.ts' WHERE id='fix-build';

4. Progress Tracking (SQL Todos)

Use todos with descriptive kebab-case IDs. Update status as you work.

INSERT INTO todos (id, title, description, status) VALUES
  ('setup-jest', 'Configure Jest', 'Install jest + ts-jest. Create jest.config.ts.', 'pending'),
  ('write-tests', 'Auth tests', 'Cover login, logout, token refresh.', 'pending');
INSERT INTO todo_deps (todo_id, depends_on) VALUES ('write-tests', 'setup-jest');
UPDATE todos SET status = 'in_progress' WHERE id = 'setup-jest';
UPDATE todos SET status = 'done' WHERE id = 'setup-jest';

5. Verbatim Recording

Record everything the user says — completely, unedited. No chunking, no summarizing, no truncating.

GOOD: Store full quote "Create 11 skill files with YAML frontmatter and practical examples"
BAD:  Store "Create skill files" — lost all critical details

6. Parallel Execution Patterns

Dispatch 5–10+ agents simultaneously. Hardware supports 100+ concurrent agents.

task(explore): "Find all TypeScript configs"       # all four launch simultaneously
task(explore): "List all test files and frameworks" # results collected as each
task(explore): "Map dependency graph of src/core/"  # finishes — no sequential
task(explore): "Find all API endpoint definitions"  # bottleneck

7. Fleet Deployment

Launch large agent batches for bulk operations. Track completion, handle partial failures.

for skill in [memory, routing, fleet, recovery, ...]:
  task(general-purpose, background): "Create SKILL.md for {skill}"
# 9 done, 1 in_progress, 1 failed → re-dispatch failed with alt model

8. 503 Retry Patterns

When rate limits hit: log failure, continue other tasks, retry with backoff (30s → 60s → 120s),
swap models if needed (opus → sonnet → haiku), queue if all models fail.

9. Proactive Backup Agents

For critical tasks, launch primary + backup on a faster model. First to finish wins.

task(general-purpose, model="claude-opus-4.6"):   "Generate arch doc"  # primary
task(general-purpose, model="claude-sonnet-4.5"): "Generate arch doc"  # backup
# 2x cost, guaranteed fast delivery. Reserve for critical-path items only.

Anti-Patterns

Anti-Pattern	Correct Approach
Asking permission for every action	Act first, report results
Sequential independent tasks	Dispatch agents simultaneously
Giving up after first failure	Retry with different approach/model
No progress tracking	Use todos table with status updates
Summarizing user input	Store complete text verbatim
Retrying same failing command	Change approach, escalate, try alt tools
Idling on blocked task	Work on other ready tasks in parallel

Checklist

• Autonomous execution verified — no unnecessary permission requests
• Progress tracked in SQL — all tasks in todos with current status
• Parallel dispatch used — independent tasks launched simultaneously
• Error recovery tested — retry-with-different-approach occurred
• Verbatim recording confirmed — user input stored complete
• Fleet deployment, 503 handling, and backup agents operational