Concept · Agent design

Agent autonomy levels, from L0 to L5

Not every task needs a fully autonomous agent. A simple ladder from L0 (a person does everything) to L5 (the agent acts unsupervised) helps you pick the right level per workflow.

TL;DR: Agent autonomy is a dial with six rungs, from L0 (a person does the task) to L5 (the agent acts unsupervised), and you set it per workflow based on risk, reversibility, and what the evals prove.

Book an AI audit

Last updated: 2026-08-07

Autonomy is a dial, not a switch

The most common mistake is treating AI as all-or-nothing: either a person does the work or the agent does it alone. In practice, autonomy is a dial you set per task based on risk, reversibility, and volume.

  • Turn the autonomy level up only when evidence justifies it.
  • Different workflows can sit at different levels in the same company.
  • Starting low and moving up is safer than starting high and pulling back.

The levels, L0 to L5

A simple ladder gives a shared language for deciding how much control to delegate. The names matter less than the question: who decides, and who approves?

  • L0 — A person does the task; AI does not step in.
  • L1 — AI suggests; the person decides and executes.
  • L2 — AI drafts an action; the person approves before it runs.
  • L3 — AI executes within limits; the person reviews by sampling and exception.
  • L4 — AI executes and only escalates uncertain cases to a person.
  • L5 — AI acts without routine supervision, with audit after the fact.

Choosing the level for a workflow

The right level depends on what happens if the agent is wrong. The higher the risk or the lower the reversibility, the lower on the ladder you start.

  • Is the action reversible? If not, keep human approval.
  • What is the cost of an error? High cost, lower level.
  • Is there enough volume to make automation worth it?
  • Do evals show stable accuracy on this workflow?

Questions buyers ask

Should every workflow reach L5?

No. Many high-value workflows live comfortably at L2 or L3. The goal is not maximum autonomy — it is the highest level that is still safe and useful.

Can the level change over time?

Yes, and it should. A workflow can start at L2 and move to L4 once reliability metrics support it.

AI operator field notes

illmethinks.io publishes source-transparent notes on AI agents, tools, and operational risk monitored by Paput.ai.