Levels of Autonomy (AI Agents)
A five-level scale for classifying how much of a task an AI agent runs without human intervention—the agent analog of SAE's self-driving levels, and the antidote to Agent Washing. **L0 Scripted:** fixed rules / conversation trees (a chatbot or RPA flow). **L1 Assisted:** suggests and drafts, human executes every action (a copilot). **L2 Supervised:** plans and acts on routine steps, human approves high-stakes ones via an Approval Gate. **L3 Conditional:** runs end-to-end in a bounded domain with a Human-on-the-Loop (HOTL) monitoring. **L4 High:** runs open-ended tasks and self-corrects, human sets goals and guardrails. The honest floor for calling a product an 'agent' is L2; most real production agents live at L2–L3, and for anything touching money, contracts, or customers that supervised range is the responsible design, not a limitation.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
In a support deployment: an L1 copilot drafts replies for an agent to send; an L2 agent resolves password resets autonomously but routes refund requests to an Approval Gate; an L3 agent handles all tier-1 tickets end-to-end while a supervisor watches the queue for exceptions. Writing the target level into the contract ('operates at L3 for tier-1 tickets, measured over a 30-day pilot') makes the claim testable in a way 'agentic' never is.
Frequently asked questions
- What autonomy level should I buy?
- Match the level to the job, both directions. Deterministic, stable-rules work → L0 workflow (don't overpay for autonomy you won't use). A human should keep final judgment → L1 copilot. Multi-step, ambiguous work needing runtime decisions and recovery → a real L2–L3 agent. Overpaying for an unusable L4 and buying L0 for an L3 problem both waste money.
- Is fully autonomous (L4) better?
- Not for most business tasks. Full autonomy fits low-stakes, reversible work. For high-stakes or irreversible actions, Human-in-the-Loop (HITL) supervision at L2–L3 is the standard—a vendor promising unsupervised L4 for your accounts-payable run is either agent washing or being reckless.