Every AI employee's Agent Operating Procedure defines what it can resolve on its own and where its authority ends. Needs Attention is what happens at that boundary: instead of forcing a resolution it isn't confident in, or silently dropping the case, the agent flags it, attaches everything it has already gathered, diagnostic detail, the specific rule that was triggered, the data it drew from, and routes it to the person who owns that decision.

This is different from a generic error state or a support ticket. A Needs Attention flag always carries a specific reason, a price variance beyond tolerance, a contract clause that doesn't match the standard playbook, a ticket touching a safety system, so the person reviewing it starts with the answer to "why was this flagged" already in hand, rather than having to reconstruct it.

Frequently Asked Questions

What triggers a Needs Attention flag?

Whatever the role's Agent Operating Procedure defines as outside scope: a threshold breach, a case type marked as always-escalate regardless of how routine it looks, missing or contradictory source data, or a pattern the agent hasn't seen a reviewed resolution for before.

How is this different from an agent just failing or timing out?

A failure is unplanned; Needs Attention is a designed outcome. The agent recognized the case as outside its authority and handled it correctly by escalating, the same way a competent new hire asks a manager instead of guessing on something they were never trained to decide.

What context comes attached to a flagged case?

Whatever the agent already gathered while working the case: the specific data points it compared, the rule or tolerance that was triggered, and, where relevant, what a similar past case resolved to. The point is that the reviewer never starts from zero.

Does a high volume of Needs Attention flags mean something is broken?

Zamp addresses this by treating escalation rate as a tuning signal, not a failure signal. A high rate usually means the Agent Operating Procedure is too conservative or genuinely hasn't encountered a recurring case type yet. Once a person resolves it, that resolution gets folded back into the AOP, so the same situation is handled automatically next time.

Can an agent act first and flag it as Needs Attention afterward?

No, not for anything the AOP defines as requiring sign-off. Needs Attention is a hold state: the case waits for a person's decision before it proceeds, rather than the agent acting on its best guess and flagging the action for review after the fact.

Is every escalation logged?

Yes. Every Needs Attention flag, and its eventual resolution, is part of the same activity log as every other action an agent takes, so a manager can review escalation patterns across a role or a time period, not just individual cases as they come up.