Only after two kinds of evidence agree, for one capability at a time. Production records at the current depth must show the work holding up over a full business cycle, and a fresh certification run at the next depth must clear every gate on your own cases. The role's ceiling must allow it, and reviewers must be ready.
What does widening actually change?
Depth is how far an agent may act for a given capability: read, draft for a person to send, execute routine work within bounds, move money or take hard-to-undo actions, or coordinate other agents. Our entry on what depth means sets out the scale.
Widening changes who is the last check. At depth 2, a person sees every output before it leaves. At depth 3, most outputs leave without anyone seeing them, and people review samples, held items and escalations. That is a large change in risk for what looks like a one-step move, which is why it needs evidence from both directions.
What evidence should trigger the decision?
| Evidence | Where it comes from | What it should show |
|---|---|---|
| Edit rate at the current depth | Production drafts over a full cycle | Reviewers rarely change outcomes, only style |
| Scope and refusal behaviour | Filed verdicts | No attempts outside bounds; refusals where expected |
| Held and escalation patterns | The held queue | Stable rates; held items mostly approved for clear reasons |
| Certification at the next depth | A fresh run on recent cases | Every gate clears on its own |
| Drift signals | Weekly review | No movement since the last certification |
| Role ceiling | The agent's role | The next depth is within it |
Each row tests something different. A low edit rate shows the drafts are good. It does not show the agent would handle the cases that reviewers caught before they went out, which is what certification at the next depth tests. A passing certification without production evidence shows the agent can do the work on a test set, not that it does so on live volume. The step where both are checked is sometimes called a depth gate.
What should not trigger widening?
- Volume pressure. A growing queue is a staffing problem. Widening to clear it puts risk on customers to solve a rota.
- Confidence. "It has been fine for weeks" is not evidence unless it is backed by filed verdicts and scores. Our view on why widening depth is an evidence decision, not a confidence decision sets this out.
- A vendor's results from other customers' cases.
- Success in another capability. Each agent earns its depth on its own cases.
Who decides, and how?
- Analyst: assembles the production evidence for the capability over the full cycle. Evidence: a short evidence pack.
- Process owner: requests certification at the next depth on a refreshed case set. Evidence: the request and the case set version.
- Agent: runs the certification. Evidence: scores per gate.
- Team lead: confirms reviewer readiness for the new pattern of work, such as sampling instead of sending. Evidence: an updated rota and sampling plan.
- Process owner, with risk for depth 4: approves and sets a review date. Evidence: a filed decision naming the evidence.
What does a good widening decision look like?
Say a bank's mortgage servicing agent has drafted replies about overpayments at depth 2 for three months. Reviewers changed outcomes in very few drafts, all involving early repayment charges on older products.
The owner asks to widen overpayment replies to depth 3, but only for current products, excluding the older ones where corrections clustered. The certification on recent cases for current products clears every gate. Reviewers move from sending every reply to sampling a set share each day and handling held items. The older products stay at depth 2. This is the pattern described in widening an agent's depth one capability at a time: narrow, evidenced and reversible.
Narrowing works the other way, and much faster; see when an AI agent's depth should be reduced.
The wider sequence is in our guide to deploying an AI agent.
Where Pryme fits
In Pryme, depth is earned, not configured. The role sets a ceiling, Attest sets the depth from evidence on your cases, and the lower wins. A certificate lapses after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, so widened depth has to keep being earned.
Questions people ask
How long should an agent stay at a depth before widening?
At least one full business cycle for that capability, including its peak. For many processes that means a month or two, and longer for capabilities that move money.
Can an agent skip a depth?
It should not. Each depth produces the evidence for the next. An agent that goes from read-only to executing has no production record of drafting to support the move.
Who can approve depth 4?
The process owner with a risk sign-off, because depth 4 covers moving money and actions that are hard to undo. Some organisations add a second senior approver.