Set a new agent's first limits by asking what you could afford to get wrong before anyone noticed. Cap depth one level below what the job eventually needs, cap value at what a new junior could approve, and cap daily volume at what a reviewer can check in full. Revisit after week one, any incident and each renewal.
What is a ceiling, and why does a new agent need one?
A ceiling is the most an agent may do, whatever its evidence says. It has three parts: the highest depth each capability may reach, the largest value any single action or day's actions may carry, and the most actions it may take in a period. It sits in the agent's bounds as the customer's own limits, one of the five bounds in the minimum chain, and like the others it wins when it is the lowest.
The distinction from depth matters. Depth is earned from evidence: certification on your cases. The ceiling is set by you, and it can be lower than the evidence supports. In our model, the agent's role sets an outer limit, evidence sets depth, and the lower wins; the autonomy ceiling and customer ceiling explain the two layers.
A new agent has certification but no live history. Certification shows it handles past cases; it does not show how it handles today's surprises in your live systems. The first ceiling covers that gap.
What should the first ceiling be?
The principle: the ceiling should be small enough that the worst plausible day is something you would catch and could fix.
Depth. Start each capability one level below where it will eventually sit, or at the certified depth if that is already 2. If the agent is certified to depth 3, a capability that needs 3 starts at depth 2 for its first days, then rises when the live drafts match certification. Anything that moves money or is hard to undo starts held, whatever the certificate says.
Value. Cap single actions at what you would let a capable new joiner approve alone in their first month. Cap the daily total at an amount that, if every action that day were wrong, you could reverse or absorb. If you cannot name that amount, the capability is not ready to act on value.
Volume. Cap the daily count at what one reviewer can check in full after the fact. That way, for the first days, every action is seen. As the reviewer moves to sampling, volume can rise.
Scope. Limit case types, channels and hours. Staffed hours only. One queue. The case types that certification covered most thoroughly.
What are the steps?
- Agent owner: lists capabilities and the depth each will eventually need.
- Agent owner and risk: set the first depth for each, using the rule above. Evidence: the depth table, signed.
- Agent owner: sets value caps, single and daily, per capability. Evidence: the caps with the reasoning.
- Agent owner: sets the volume cap from reviewer capacity. Evidence: reviewer capacity and the cap.
- Agent owner: sets the scope: case types, channels, hours.
- Agent operations: loads the ceiling into the bounds and tests that an action over each cap is held or refused. Evidence: test results showing the deciding bound.
- Agent, at ceiling depth: works, with every action passing the minimum chain and the verdict naming which bound decided.
- Agent owner: reviews at the set points below, and records any change.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| List capabilities | Nothing | none | Agent owner |
| First depth | Nothing | none | Owner and risk |
| Value caps | Nothing | none | Setting them |
| Volume cap | Nothing | none | Setting it |
| Scope | Nothing | none | Setting it |
| Load and test | Routes test actions | 1 | Checking verdicts |
| Work | Acts within ceiling | up to ceiling | Held items |
| Review | Reports actions near caps | 1 | Deciding changes |
Which first process makes this easier?
The easier the work, the easier the ceiling. Routine, reversible, high-volume work with clear rules lets you set a meaningful ceiling and still learn quickly. Work that is rare, high-value and irreversible gives you a choice between a ceiling so low the agent does nothing and one so high a single error hurts. That is one reason your first agent should do boring work. The criteria for picking that work are in which process you should give an AI agent first, and the method in how to choose the first process.
When should the ceiling be revisited?
After the first week. If every action was reviewed and the work was clean, raise volume and allow depth to reach the certified level.
When the agent routinely hits a cap. Frequent held or refused actions at the volume or value cap mean the cap is shaping the work. That can be right, but it should be a decision.
After any incident. Lower first, investigate, then decide.
At every certificate renewal, every 90 days in our model. Fresh evidence may support a higher ceiling or argue for a lower one.
When the work changes: a new product, a new case type, a change in the systems.
Raise one part at a time. Raising depth, value and volume together makes it impossible to tell which caused a problem.
Where does it usually go wrong?
The ceiling copies the certificate. The certificate allows depth 3, so the first ceiling is depth 3. The certificate is an outer limit, not a starting point.
Value caps are set per action only. A low single cap with no daily cap lets the agent take a large number of small actions. Set both.
Volume is set by system capacity. The agent could handle thousands a day, so the cap is thousands. The cap should follow what people can check.
No one revisits. The first ceiling stays for a year, and the agent is held far below what it has earned, or quietly bypassed. Put the review dates in the calendar, as part of train, shadow, certify, widen.
Questions people ask
Is the ceiling the same as depth?
No. Depth is earned by evidence; the ceiling is set by you. The lower of the two decides what the agent may do.
Can the ceiling differ by customer or region?
Yes. Ceilings can be set per segment where the risk differs, for example lower value caps for a market where reversals are harder.
Who can raise the ceiling?
The agent owner, with risk's agreement for capabilities that move money or affect customers significantly, and always recorded.