Find errors in production agent traces

Laminar's flow-1 model reads agent traces to catch implicit failures like wrong clicks or wrong citations, so every production trace can be analyzed without sampling.

What They Did

flow-1, a small open-source model fine-tuned and then trained with RL on synthetic investigations, runs inside a harness called Signals. It treats a trace like a repo with each span as a file, starts from a compressed trace plus a user prompt, greps spans, reads relevant ones in full, and returns findings as JSON in a user-defined schema.

How It Went

Laminar says it matches GPT-6-sol in detection quality on its own benchmark at 23x lower cost, making analysis of every trace affordable. Similar findings are clustered, though the benchmark is the vendor's own.

Worth Knowing

Signals accepts general prompts, such as checking whether a browser agent completed a checkout, and clusters of failures can seed eval datasets for fixing them. Laminar is Apache 2.0 and self-hostable.