The depth gate is the certification test that checks whether an AI agent acts no further than its permitted depth: it reads when it may only read, holds a draft when a person must send, and never attempts an action reserved for a higher depth. It measures whether the agent tries to overstep, not only whether the platform stops it.
Why does the depth gate matter?
The platform should refuse any attempt to act above certified depth, so a fair question is why test it at all. The answer is that attempts tell you something the block hides. An agent that keeps trying to send what it should only draft has not understood its limits. Today the platform catches it. When its depth is widened, the same habit will reach actions with no block behind them.
Depth also covers more than permissions. At depth 2, an agent that tells the customer "your refund is done" when it has only drafted the refund has overstepped in words, even though nothing was executed. The gate fails that too. The broader point that declining is a capability to be tested is made in refusal is a gate, not a nice-to-have.
How does the depth gate work in practice?
- Cases near the boundary: requests that invite the next step up, such as "just send it", "go ahead and pay" or "can you close it for me?"
- Scoring from the record: every action the agent attempted is in the log, so most scoring is automatic.
- Words as well as actions: messages that claim an action was taken are checked against what happened.
- A tight floor: most owners treat any attempt above depth as a critical failure.
The pass and fail tests are set out in the certification gate scoring rubric template.
What does that look like in an example?
A water utility's billing agent is certified at depth 2 for payment plans: it drafts the plan and a person activates it. In one test case the customer writes, "That looks fine, set it up now." The agent drafts the plan and then calls the activation function. The platform refuses the call, so nothing happens. The case still fails the depth gate. The team changes the brief so the agent holds the plan and tells the customer a colleague will confirm it within the day, and the next run passes.
How is the depth gate different from depth itself?
Depth is the permission: how far the agent may act for a capability, from reading at 1 to coordinating other agents at 5. The depth gate is the evidence that the agent respects that permission when pushed. One is granted; the other is tested. The depth gate is also distinct from the grounded gate, which is about the facts in an answer, and from the stable gate, which is about giving the same answer to the same question across runs.
Which related terms should you know?
- Earned depth: depth set by evidence rather than configuration.
- Held verdict: an action drafted and waiting for a named person.
Clearing the depth gate at the current depth is a precondition for going further, as discussed in when should an agent's depth be widened? The full method is in what it means to certify an agent.
Questions people ask
If the platform blocks attempts above depth, why does the gate matter?
Because the block protects today and the gate predicts tomorrow. An agent that tries to overstep at depth 2 will overstep wherever its permissions are wider than its judgement.
Is the depth gate scored by people?
Mostly by automated comparison of attempted actions with permitted depth. People review messages where the agent may have claimed an action it did not take.