5 levels of trusting an AI agent

Each level is defined by how much of the agent's work a person still checks.

Like self-driving, agent autonomy comes in levels. Every agent can write code, so the useful measure is how much of its output a person still inspects before it counts. Tap a level.

The gate shrinks but never closes

Each level up, a person inspects less: every line, then every PR, then the risky slice, then only exceptions. Even at the top, a person owns the intent, the exceptions and the outcomes. In 2026, "no humans in the loop" is a demo and a sales pitch.

Where teams are

Most serious teams sit around Level 2 to 3: agents open pull requests and people review them. Level 4 means trusting a merge you didn't read, and trust in AI-written code has been falling. Guardrails you believe in get you there. That means evals that catch the failure modes, small diffs, and a fast way to get human eyes on what machines can't judge. Each step up waits on review. If review stays narrow, you stay at Level 2 however good the agents get.

How to use it

Rate yourself per kind of change. You can be at Level 4 for dependency bumps and docs and Level 2 for anything touching auth or money. That's correct. You earn autonomy one risk category at a time, with evidence, as fast as your review keeps up. For the bigger picture, see what a software factory is.