Agent runs
See what your agent actually did
You showed it once. Your agent ran at 3am. In the morning you have a result and no idea how it got there. Record the run and you get the review: the steps it took, the ones it skipped, the moments worth checking, and one sentence on whether to trust what it produced.
Three answers, and we mean all three
Completed
It did the task, the way you showed it. Nothing to look at, which is the answer you want most mornings.
Completed, with deviations
It got there another way: a step of your flow skipped, an extra one taken, two of them out of order. Each one is a line you can read, with the second it happened at.
Failed, or unclear
It did not finish, or the recording does not say. We would rather tell you we cannot tell than guess, because a review you cannot trust is worse than none.
Against the flow you showed it
Point a run at a recording of how a person does the same job and the review compares the two, step by step: matched, skipped, reordered or extra. It lines up the two click trails, so the recording you point it at needs one of its own. An agent that skipped a step of your flow cannot be reported as having completed it, whatever else went well.
What it is not
- Not a trace viewer. It reads the recording and the click trail, not your agent’s prompts, tool calls or tokens. Your framework’s own tooling is better at that, and this is for whoever signs off rather than whoever builds.
- Not live. It reviews a run that already happened; it cannot stop one.
- Not a guarantee. It sees what the recording shows, and it says so when that is not enough.
- Typed text is never recorded, and a sensitive field produces a marker and nothing else.
The documentation has the setup, the full list of what is recorded, and the limits in detail.