semloop

At a glance

semloop checks an agent trace for steps that repeat without making progress, even when the wording changes. It is worth comparing with your existing checks if stalled runs look different on the surface but keep doing the same work.

Good fit

Teams reviewing agent traces who can label examples of both useful retries and genuine stalls.

Consider something else when

Skip it if you need to stop a live run immediately or cannot share traces with the judge you configure.

Start here

Start with two synthetic traces. In one, have the agent rephrase the same failed attempt. In the other, let it discover new evidence before retrying. Check whether the detector’s reasons distinguish them.

Cost and requirements

MIT-licensed Python project. It can use TypeSafe Jev or an OpenAI-compatible judge, with credentials and usage costs set by that provider. The mock backend supports scripted tests.

Limits to know

A judge may mistake a useful retry for a loop or miss progress that the trace fails to capture. Treat its score as a review signal and check what trace data goes to the judge.

Related work

• JevMan (context)