Graduated autonomy is the practice of increasing an AI agent's independence in defined steps as evidence of good performance builds up, and reducing it when that evidence weakens. An agent starts by reading, then drafting, then acting within narrow bounds, and moves up only when it has proven each step on real cases. It treats autonomy as something earned and kept, not configured once.
Why does graduated autonomy matter once agents act?
Because no test before launch can show everything. An agent that performs well on a sample may meet cases in production that the sample never covered. Starting low and widening on evidence limits the cost of what the tests missed, and builds a record that each widening was justified.
It also gives the business a way to say no without abandoning the agent. A capability that is not yet ready stays at draft while others move up.
How does graduated autonomy work in practice?
Each capability has its own track. The steps usually follow a depth scale: 1 read, 2 draft for a person to send, 3 execute routine work within bounds, 4 move money or take hard-to-undo actions. Moving up requires evidence from the agent's performance on real cases from the business's own operation, scored against a floor. Moving up beyond the ceiling the business has set requires a separate, recorded decision through an autonomy change request.
Graduation runs both ways. Evidence expires and must be refreshed. A failed run, an incident or a change in the agent's brief or model can drop a capability back down. Doing that well after something goes wrong is covered in the autonomy reduction plan template after an incident.
What does that look like in an example?
A building society introduces an agent for mortgage overpayment requests. In the first month it reads requests and summarises them for staff. After its summaries score well against the society's own cases, it drafts confirmations that staff send. After further evidence, it processes standard overpayments within product limits on its own. Requests that trigger early repayment charges stay at draft, because the agent's handling of charge explanations has not yet cleared the floor. Six months later its certificate is renewed on fresh cases.
How is graduated autonomy different from autonomy levels?
Autonomy levels are the scale: the named steps of independence. Graduated autonomy is the process of moving along it, with evidence required at each step and the possibility of moving down. A scale without graduation is a label someone picks. Graduation without a scale has nothing to measure progress against. How the two fit together is covered in how autonomy levels for AI agents work.
How does it relate to supervised autonomy?
Supervised autonomy is a mode in which the agent acts while a person monitors and can intervene. It is often a step on a graduated path: the agent executes, but under close watch, before supervision relaxes. Graduation decides when the agent may act; supervision decides how closely it is watched while it does. The underlying principle is in regulated autonomy.
Where Pryme fits
In Pryme, depth is earned, not configured. The agent's role sets a ceiling, Attest evidence sets the depth, and the lower wins. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.
Questions people ask
How long should an agent stay at each step?
Until the evidence clears the floor on enough real cases, not for a fixed period. Time alone proves nothing; a quiet month may not include the hard cases.
Does a model update reset graduated autonomy?
It should trigger fresh evidence. An agent on a new model is, in effect, a different agent, so its capabilities should be re-certified before keeping their depth.