Human-in-the-Loop Governance

Autonomy is a dial, not a switch. The harness controls where it's set.

The autonomy spectrum

Not all agentic tasks warrant the same level of human oversight. A well-designed Agentic SDLC defines explicit autonomy levels and assigns tasks to them deliberately.

Level	Agent autonomy	Human involvement
L0 — Assisted	Suggests changes; human applies	High
L1 — Supervised	Applies changes; human reviews each	Medium
L2 — Checkpointed	Runs a session; human reviews at end	Low
L3 — Autonomous	Runs multi-session without review	Minimal

Most production teams operate at L1–L2. L3 is appropriate only for well-bounded, high-coverage tasks.

When to insert a human checkpoint

The harness defines explicit conditions that pause autonomous execution and request human review:

• Scope boundary hit — agent attempted to modify an out-of-scope file
• Verification failure — tests failed more than N times in a row
• Ambiguity detected — task description conflicts with AGENTS.md rules
• Architecture impact — proposed change touches a core interface or schema
• Novel failure mode — error type not seen in previous sessions

The review protocol

When a human checkpoint is triggered, the agent writes a structured review request:

# Review Request — 2026-04-17

## Task
Implement user authentication logout endpoint

## What I did
- Added `POST /auth/logout` endpoint in `src/auth/routes.ts`
- Invalidated session token in Redis

## Why I stopped
Verification failed: `tests/auth/logout.test.ts` — 2 failing assertions

## What I need
Clarification on expected token invalidation behaviour (see failing test line 47)

## Next step if approved
Fix token invalidation logic and re-run verification

The human reads, responds, and the agent resumes. The entire exchange is committed to git.

---

Next: 09. Bibliography →