Earned depth: what it means for AI agents

Earned depth is the principle that an AI agent's depth, how far it may act, is set by evidence of how it performs on the customer's own cases rather than by a configuration setting. Certification evidence sets the depth, up to the role's ceiling, and no plan limits it.

Why does earned depth matter?

Configured autonomy is trust by declaration. Someone ticks a box, and the agent may now send payments. If asked why, the honest answer is often that it seemed fine in testing. Earned depth replaces that with trust by evidence: the agent may send payments because it was replayed against hundreds of the business's own past cases and cleared every gate at that depth.

It also works in reverse. Depth that is earned can be lost. In our model a failed run cuts the agent to the highest depth it cleared, or to depth 1 if it cleared none, the same day, the position argued in a failed run should cost depth the same day. A setting would stay where it was until someone remembered to change it.

How does earned depth work in practice?

Depth runs from 1 to 5: read, draft for a person to send, execute routine work within bounds, move money or take actions hard to undo, and coordinate other agents. For each agent:

1. The owner requests a depth.
2. The agent is certified at that depth on the customer's cases, and every gate must clear, including the depth gate.
3. The evidence supports a depth, which may be lower than requested.
4. The effective depth is the lower of that and the role's ceiling.
5. Monitoring continues; depth widens on new evidence and narrows on failure or drift.

The conditions for each direction are covered in when should an agent's depth be widened? and when should an AI agent's depth be reduced?

What does that look like in an example?

A wealth manager's client service agent has a role ceiling of depth 3. The owner requests depth 3 for updating client contact details and for applying small fee rebates. The agent handles contact details well, but on fee rebates it misses several cases where a rebate needed a manager's sign-off, so the evidence supports only depth 2: it drafts both, and a person applies them. Two months later, after the brief is clarified, a new run supports depth 3.

How is earned depth different from an autonomy ceiling?

An autonomy ceiling is the most an agent could ever be allowed under its role or the customer's policy. It is configured, and it caps everything. Earned depth is what the agent has actually proven, below or at that ceiling. A high ceiling grants nothing on its own; an agent with a ceiling of 4 and evidence for 2 acts at 2. A customer can always set limits lower than the agent has earned. Nobody can set them higher.

Where Pryme fits

In Pryme, evidence from Attest sets the agent's depth, up to its role's ceiling, and no plan limits it. A run that clears no depth drops the agent to depth 1 the same day, and certificates lapse after 90 days, so depth is earned again rather than carried forward.