Autonomy Levels Matrix
The Five Levels
| Level |
Name |
Agent Acts? |
Human Reviews? |
Human Approves? |
| L0 |
Manual |
❌ |
N/A |
N/A |
| L1 |
Assisted |
Suggests only |
✅ Every action |
✅ Every action |
| L2 |
Supervised |
✅ Yes |
✅ Every output |
✅ Before publish |
| L3 |
Monitored |
✅ Yes |
Exceptions only |
❌ Not required |
| L4 |
Autonomous |
✅ Yes |
Periodic audits |
❌ Not required |
Task Classification Criteria
When classifying a task, answer three questions:
| Question |
Low Score → Higher Level OK |
High Score → Lower Level Required |
| Reversibility |
Easily undoable (delete a comment) |
Hard to undo (deploy to prod) |
| Blast Radius |
Affects 1 file in 1 PR |
Affects security config or user data |
| Predictability |
Same task done correctly 50+ times |
Novel task or unstable codebase |
Level Guidance by Task Type
| Task Type |
Suggested Level |
| Fix typo in docs |
L3 |
| Add unit tests for existing function |
L2–L3 |
| Implement new API endpoint |
L2 |
| Refactor authentication module |
L1–L2 |
| Modify CODEOWNERS / security config |
L1 |
| Create/delete repository |
L1 |
| Infrastructure changes |
L1 |
Guardrail Requirements by Level
| Level |
Required Guardrails |
| L1 |
N/A (human approves every action) |
| L2 |
CODEOWNERS on sensitive paths, PR review required |
| L3 |
Environment approval gate for deployments, audit logging |
| L4 |
All of L3 + automated anomaly detection + kill switch |
Q18: Autonomy Levels Matrix — Full implementation with _data/autonomy-matrix.yml schema.
Part of: [[GH-600 Agentic AI Quick-Reference Notes]] · Related quest: [[The Autonomy Scales: Mapping Agent Autonomy Levels]] · Hub: [[The Agentic Codex: GH-600 Study Hub]]