How to widen an AI agent's depth one capability at a time

Choose one capability, gather the evidence that it is ready, and certify the agent again at the new depth before anything changes. Then adjust its ceiling, approver roster and fallback plan, switch on a staffed day, and review every action at the new depth for the first week. Keep every other capability unchanged until this one settles.

Why only one capability at a time?

An agent does several jobs, and each job needs its own depth. Reading a balance, drafting a letter and issuing a refund are separate capabilities with separate risks, which is why each action needs its own depth. Widening them together might seem efficient, but if something goes wrong, you cannot tell which change caused it, and you have to pull them all back.

One at a time also keeps the evidence honest. Each widening is a small, testable claim: this capability, for this customer, on these cases, is ready to go one level further. The capability and depth matrix template is a useful place to record each capability's current and proposed depth.

The agent owner proposes and decides. The certification lead certifies. Risk reviews any widening to depth 4, where the agent moves money or takes actions that are hard to undo.

What are the steps, and what evidence does each leave?

1. Agent owner: proposes the widening for one capability, with the reason and the evidence from its current depth: reviewer edit rates, the share of held items approved unchanged, random sample results, and incidents. The signals that justify it are covered in when an agent's depth should be widened. Evidence: the proposal.
2. Agent owner: checks the ceiling. The role's ceiling and the customer's own limits must allow the new depth. Evidence: the ceiling check.
3. Certification lead: certifies at the new depth. The replay must include cases the current depth caught: drafts that people corrected, items that were rightly held. At the new depth, the agent would act on these alone. Evidence: the certificate, with every gate's result.
4. Risk: reviews if the new depth is 4. Evidence: the review and its conditions.
5. Agent owner: sets the new limits: value and volume caps for the widened capability, starting lower than the eventual level. Evidence: the limits in the bounds.
6. Agent owner: updates the approver roster and fallback plan. Fewer items will be held; approvers' work changes. The fallback team must still be able to take the work back. Evidence: updated roster and plan.
7. Agent owner: decides what the freed time will be used for before it is freed. Evidence: the decision.
8. Agent operations: changes the depth on a staffed day. Evidence: the time the change took effect.
9. Reviewer: checks every action at the new depth for the first week, then moves to sampling. Evidence: the review log.
10. Agent owner: confirms or reverses at the end of the first week. Evidence: the decision.

Why must certification include the cases people caught?

At the current depth, people catch the agent's errors before they matter. Those are exactly the cases that would get through at the new depth. If the recertification replays only clean cases, it will pass, and the first real error will reach a customer. Include every case from the last period where a reviewer made a substantive edit or an approver rejected, and require the agent to handle them correctly or hold them.

Why decide about freed time first?

Because widening depth frees people's time, and that time will be used for something whether or not anyone decides. Often it disappears into other work, and the benefit cannot be seen or measured. Sometimes it creates anxiety about jobs. Deciding in advance, whether to redeploy people to exceptions, reduce a backlog or slow hiring, makes the change visible and fair. The argument is in decide what to do with the time agents free up before they free it.

What should you measure after widening?

• Errors at the new depth, from full review in week one and sampling afterwards.
• Refusals and holds at the new limits: are the caps doing their job?
• Customer contacts and complaints about the widened work.
• Downstream rework.
• Approver workload, which should fall.

When should it go back down?

If the first week finds errors that matter, reverse the widening at once, before investigating. A failed drift check or an incident does the same automatically. The conditions for lowering depth are set out in when an AI agent's depth should be reduced. Reducing depth is not a failure of the process; it is the process working.

Where does it usually go wrong?

Several capabilities move together because they were all ready. Stagger them, even by a week.

Recertification reuses the old case set. The cases that proved depth 2 do not prove depth 3. Add the cases people caught.

Limits jump to the eventual level. Start the widened capability below its final caps and raise them as evidence grows.

Approvers are not told. They notice fewer held items and assume something is broken, or they keep checking in the old way and duplicate effort.

The widening is never reviewed. Week one passes without a decision. Make confirmation an explicit step, as in train, shadow, certify, widen.

Where Pryme fits

In Pryme, depth is held per agent, and a widening takes effect only when Attest has issued a certificate at the new depth and the role's ceiling allows it. The lower of the two always wins. Every action at the new depth is filed with the bound that decided it, so the first week's review works from a complete record.