How is autonomy earned?#
One class of work at a time. An agent that has written reliable documentation for months has proved nothing about its database migrations. Autonomy usually widens in order of consequence: documentation, then tests, then implementation, then production paths. Each class moves once the evidence behind its check has held up for that kind of change. Human-factors research calls the goal calibrated trust: reliance that matches what the automation can actually do (Lee and See, 2004).
Why start with more review than you need?#
Because a gate removed without evidence is a guess. Deliberate over-supervision produces the data that justifies each later removal, so the delegation boundary moves for a reason someone can write down. AI agent evaluation supplies much of that evidence.
Skip this step and reviews arrive faster than anyone can read them, get approved out of habit, and leave a record of human oversight that never happened.