Human Guides / Autonomous Mode

🤖 Autonomous Mode

What actually happens when you step away. Agents build, test with MCPs, the janitor keeps .ay/ honest, and you come back to results — not questions.

Agents self-test with MCPs Janitor keeps .ay/ accurate You review once, approve once
The full picture
What happens while you're away — minute by minute
You brief the operator, step away. Here's what the system does on its own.
T+0:00
You
Tell the operator what to build. Answer plan-phase questions (framework, DB, auth). Say "go, I'll be back in 2 hours."
your only input until review
T+0:02
Operator
Spawns 4 agents — each in its own git worktree + cmux terminal. Each gets a task spec with all decisions pre-loaded. No agent needs to ask you anything.
4 Claude processes running
T+0:05
All Agents
Run self-inventory: check tools, context budget, constraints. Claim lock files. Start building in their scoped directories. Spawn their own subagents for parallel subtasks.
parallel builds underway
T+0:30
Operator (fleet-scan tick)
Scans all terminals. collab-agent BLOCKED — needs WebSocket decision. Checks DECISIONS.md → y-websocket was decided in plan phase → sends answer to collab-agent. No need to wake you.
T+1:10
Drift Detected
canvas-agent reports done. Operator runs drift check: git log main..HEAD --oneline → empty. Zero commits. Agent described work but didn't build it.

Operator spawns verifier → confirms nothing exists → re-dispatches with explicit: "only code counts."
caught automatically
T+1:45
Janitor Agent
Operator spawns janitor after 3 tasks complete. Checks each: export-agent has commits + merged PR → real done ✓. canvas-agent had zero commits when marked done → reverts to IN PROGRESS on BOARD.md.
BOARD.md corrected
T+2:00
canvas-agent (second attempt)
Real build this time. Opens Playwright MCP → navigates to localhost:3000/editor → verifies canvas renders, pan works, zoom works. All checks pass. Commits with test evidence.
commits + test evidence
T+2:15
Operator
All tasks done. Final janitor pass → BOARD.md accurate. Creates PRs. Writes session summary to .ay/tracking/REPORT.md.
T+2:20
You return
Read cat .ay/tracking/REPORT.md. 4 PRs ready. Open each, check test evidence. Merge what looks good.
your review: ~15 minutes
What you come back to
Test reports with real evidence
Not "I think it works" — actual output from actual tests run by the agent.
Task 03 — Canvas + pan/zoom — DONE
3/3 tests passed
canvas-agent · agent/canvas · 3 commits
✓ Canvas renders at /editor
✓ Pan (spacebar+drag): viewport offset changed from 0,0 to 150,80
✓ Zoom (ctrl+wheel): scale changed from 1.0 to 1.4 after 3 scroll ticks
HANDOFF to toolbar-agent: Use fabric object.id not index. Grid is layer 0 — do not reorder.
Task 02 — Auth — BLOCKED
Needs your input
✗ .env.local has placeholder SUPABASE credentials
What you need to do:
Go to app.supabase.com → Settings → API → copy URL and anon key into .env.local. Then tell operator to resume task 02.
While waiting: agent started task 03 and 07 which have no credential dependency.
When you come back
Your 3-command return ritual
# 1. What happened while I was away?
cat .ay/tracking/REPORT.md

# 2. What's ready to review?
gh pr list

# 3. What needs me? (credentials, decisions)
cat .ay/tracking/BLOCKERS.md
The deal: You answer questions at the start (plan phase). You handle blockers needing real credentials or decisions. You review PRs at the end. Total active time: 20–30 minutes per major feature set. Everything else is autonomous.