Depth is how far an AI agent may act for a single capability, on a scale of five: 1 read, 2 draft for a person to send, 3 execute routine work within bounds, 4 move money or take actions that are hard to undo, and 5 coordinate other agents. One agent can hold different depths for different capabilities.
Why does depth matter?
Most discussions of agent autonomy treat it as one setting for the whole agent: supervised or autonomous. Real jobs do not work that way. A support agent may be trusted to update an address but not to issue a refund. A finance agent may match invoices freely but must never release a payment. Depth lets the business say exactly that, capability by capability.
It also gives a shared language. "Depth 2 for refunds" means the same thing to the operations lead, the risk team and the auditor: the agent drafts, a person sends. That precision is part of why agent governance is a bound, not a model.
How does depth work in practice?
| Depth | The agent may | Example |
|---|---|---|
| 1 Read | Look up and answer | Tell a customer their balance |
| 2 Draft | Prepare an action a person sends | Draft a complaint response |
| 3 Execute | Carry out routine work within bounds | Update a delivery address |
| 4 Move money or act irreversibly | Make payments or hard-to-undo changes | Issue a refund under the limit |
| 5 Coordinate | Direct other agents | An orchestrator assigning work |
Depth is earned, not configured. The agent's role sets a ceiling; evidence from certification on the customer's own cases sets the depth; the lower wins. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 for that capability the same day. Working out the right depth for each capability is covered in how to assign depth to each of an AI agent's capabilities, and a working layout is the capability and depth matrix template for an AI agent.
What does that look like in an example?
A wealth manager's client service agent has four capabilities. It answers questions about holdings at depth 1. It drafts responses to complaints at depth 2. It updates contact details at depth 3. It has been tested for processing withdrawals, but the firm's ceiling keeps withdrawals at depth 2, so it drafts them for an adviser. One agent, four depths, each recorded, each with its own certificate and expiry.
How is depth different from an autonomy level?
Published autonomy levels usually describe a whole agent or a whole system, often as a vendor's claim about what its product can do. Depth describes one capability of one agent for one customer, and it is earned from evidence on that customer's cases. An agent described as "highly autonomous" might still sit at depth 1 for a capability it has not been tested on.
Depth is also not the same as the depth gate. Depth is the permission. The depth gate is the certification test that checks whether the agent stays within that permission when pushed, and earned depth is the principle that evidence, not a setting, decides the level.
When should depth go down?
Depth falls automatically when a certificate lapses or a certification run fails. It should also fall when something changes underneath the agent: a new model, a new data source, a change in the process, or a run of corrections by reviewers. The triggers are discussed in when should an AI agent's depth be reduced?
Questions people ask
Can an agent hold depth 5 for every capability?
No. Depth 5 is for orchestrators that coordinate other agents, and only for the coordination capability. Each agent an orchestrator directs keeps its own depth for each task, and the orchestrator cannot grant more than those agents have earned.
Does depth 4 mean the agent can move any amount of money?
No. Depth 4 means the capability may move money at all. Every action is still checked against the customer's own limits, such as amount per payment or per day, and anything above them is held for a person.