Train the team on what changes before the agent changes. Moving an agent up a depth means people stop seeing some actions before they happen. Before that move, every reviewer and approver should know the new bounds, what they will no longer review, what evidence justified the move, and how to drop the agent back the same day.
What actually changes when an agent moves up a depth?
Depth is how far an agent may act, per capability. At depth 2 the agent drafts and a person sends. At depth 3 it executes routine work within bounds. At depth 4 it moves money or takes actions that are hard to undo. Each step removes a person from the moment of action for that capability.
That is the point of raising depth, and it is also the risk. The team's job shifts from approving each action to watching patterns in the record and reacting fast when something looks wrong. People who were good approvers are not automatically good monitors. The idea that depth is earned through evidence, not switched on, is set out in training is how an AI agent earns its next depth.
What must the team know before the move?
- Which capability is moving, and only that one. Depth is per capability. An agent may move to depth 3 for issuing standard refunds while staying at depth 2 for goodwill credits.
- The new bounds. What limit applies now, which sources must be present, and what will still be held.
- The evidence. What certification showed on the business's own cases, and which gates it cleared: grounded, refuses, depth, escalates, scope, adversarial and stable.
- What they will no longer see. Actions that used to arrive as held items will now appear only in the record.
- How to watch. Which patterns in the record should trigger a look, and who reviews a sample each week.
- How to step back. Who can lower the agent's depth, and the fact that a certification run that clears no depth drops it to depth 1 the same day.
Items 4 and 5 are the ones teams skip. A reviewer who does not know which actions have left their queue will not notice when those actions start going wrong.
What does the agent do at each depth, and what stays with people?
| Depth | The agent | The team |
|---|---|---|
| 2, draft | Prepares the refund, a person sends it | Approves or declines each one with a reason |
| 3, execute within bounds | Issues standard refunds up to the limit | Samples the record weekly, handles held items above the limit |
| 4, hard to undo | Not sought for this capability | Not applicable |
Writing the table for your own agent is a useful exercise in itself. If the team cannot fill in the right-hand column, it is not ready.
What does a safe raise look like in practice?
Take an online retailer's support team whose refund agent has spent two months at depth 2. Reviewers have approved most of its drafts, and certification on the retailer's own cases supports depth 3 for refunds up to 100 on orders with a sealed delivery record.
Before the move, the team lead runs a one-hour session. Reviewers see the certification results and the new limit. They learn that refunds above 100, or on orders with only a declared delivery claim, will still be held. They agree a weekly sample of 30 executed refunds, reviewed by two named people. They practise lowering the agent's depth in a sandbox. The move happens the following Monday, and the first weekly sample is booked before it does.
Banks follow the same pattern with tighter limits; see how banks use AI agents safely. Finance teams can map their capabilities first using a finance agent depth matrix.
What controls keep the raise safe?
- A named owner for each capability, who signs off the move and any step back.
- A written sample plan for the record, starting before the move.
- Approval rights tied to readiness. Some businesses make the right to approve a depth change depend on a current certificate held by the approver, depth by depth.
- Certificate expiry on the calendar. An agent's certificate lapses after 90 days; the team should know the date.
The wider sequence for training people around agents is in training teams for AI agents, and what each depth means is in depth: what it means for AI agents.
Questions people ask
Can a team raise depth for all capabilities at once?
It should not. Raising one capability at a time keeps the evidence clear and the team's monitoring load manageable.
What if reviewers disagree with the raise?
Listen to the reasons. If reviewers have been declining a pattern of drafts, the certification cases may not reflect the real queue, and the evidence should be checked before the move.
Does the team need retraining when depth goes back down?
A short briefing is enough: held items will return, and the queue will grow. Make sure the review rota can absorb it.