Every time a custom agent runs, the result is recorded with an outcome. There are seven main outcomes, and they're kept apart on purpose. Telling "the agent handed off to a person" apart from "the agent broke" is the difference between a design that works and a bug to chase.

The agent did the job and finished. This is what you want to see most of the time.
Resolved means the agent believed it was done. It doesn't tell you whether the customer was happy. Read a few resolved runs now and then. An agent that confidently answers the wrong question also counts as resolved.
When the run belongs to a conversation, a successful finish is recorded in more detail: answered and waiting for the customer's reply, closed, or a feature request filed or subscribed to. The AI agents report counts all of these as resolved. A run that created a coding task, or that is still waiting on a code investigation, isn't counted as resolved.
The agent handed the work to a person or another agent and stepped out. That's a success: an agent that knows its limits is working as intended.
A rising escalation rate is still worth reading. Only the runs themselves show whether the agent is too cautious or keeps meeting cases nobody prepared it for.
The run is paused. It isn't over. The agent asked a teammate something and carries on once they answer.
Watch these closely. A run that waits for days is a question nobody noticed, and there's usually a customer waiting at the other end.
The run ended without finishing and without being handed on. Usually someone stopped it, or what it was working on stopped mattering, such as a conversation that ended on its own.
Something went wrong: a tool returned an error, an endpoint didn't respond, a reply was malformed. Open the run and look at the last tool call. That's where the cause usually is.
The usual suspects are an expired credential, an endpoint that changed shape, and a required value that arrived empty.
The run hit its cost limit and was stopped. It's kept apart from Failed because the fix is different. Either Max cost per run (USD) is too low for the work, or the agent is doing more than the task needs. You change the limit on the agent's Settings → Model page.
One of these now and then is normal. A pattern of them means something is looping.
The agent used up the steps allowed for one run without finishing. The limit is Max steps per run on the same Model page. It's almost always a task that's too broad, or instructions that never say when the agent is done.
Adding a clear stopping condition to the instructions fixes most of these: what "done" looks like, and what to do when the answer can't be found.
These outcomes describe a single agent run. The AI quality review scores whole conversations with its own labels.