An agent's calls keep failing
What you're seeing
A finding like:
Refund bot has had its last 4 calls in a row fail, starting 2026-11-13T09:41:12Z: issue_refund, lookup_order. Error classes recorded: http_503. The records do not say why; that is for the owner to find.
It is a Calls failing in a row finding, one of the early warnings, medium severity.
What it means
In one trace (or one session, when there is no trace), the agent's most recent calls ended in failure, three or more in a row, and the first of them was in the last 30 minutes. Calls with no recorded outcome are not counted either way. A model call that failed counts like any other.
It catches the fast case that error rate rising cannot: that finding compares a day with the fourteen days before it and needs a sample, so it is slow to notice calls that start failing back to back.
What it does not say is why. The tools and the error classes are what the records carry.
What to do
- Open the first failure and the latest. Look at the tool, the target and the error class.
- Check what changed just before the first failure: a credential that expired, a dependency that went down, a new input, a deployment.
- If the agent is stuck in a loop on the same call, you will see a repeated call finding instead: five or more of the same call is reported there, and not twice.
- The finding closes by itself when a later call in that run succeeds. Resolve it with the cause if you want it on the record.
What this does not do
It does not stop anything or judge the agent. It reports that the latest calls failed, how many, since when, and with what.