Set the limits per feature.
A research agent is meant to take many steps. An extraction pipeline isn’t. Raise a limit, lower it or switch a check off for the features where it doesn’t fit, rather than living with the noise.
Sam reads the two reports your code sends around each AI call: measurements only, never the content. These nine signals work from a feature’s first report. Every finding shows what Sam saw, and every limit can be changed or switched off per feature.
See the signalsThese read the measurements in each report, so they work from the moment a feature is connected.
Which models each feature calls, how often, and how that changes over time.
A feature still calling a model whose provider has announced a retirement date.
A call that starts and never reports back: no reply, no error, nothing that would trip an alert.
The same error coming back again and again in one run, with nothing handling it.
The context a run sends its model growing several times over, so each call does more work than the last.
These look at the run in progress. The limit on each card is the default; you can set your own for any feature.
The same call, with the same fingerprint, made again and again in one run.
The same tool called with the same arguments, over and over.
One run starting more and more child runs.
A run going past the number of steps you expect from that feature.
A research agent is meant to take many steps. An extraction pipeline isn’t. Raise a limit, lower it or switch a check off for the features where it doesn’t fit, rather than living with the noise.
Each finding shows the counts, the limit and the run it came from. If one is wrong, mark it, and Sam records it alongside the finding.
New teams start with Sam observing. Switch halting on and a finding also returns a halt to your agent on its next call. Your code decides what to do with it.
Two reports around each call. Your first findings arrive while you watch.
Start a free trial