Laminar releases flow-1 to catch silent AI agent failures

Laminar says the RL-trained model matches GPT-6-sol detection quality at 23 times lower cost on its benchmark.

Y Combinator S24 startup Laminar announced flow-1, a small open-source model post-trained with reinforcement learning to find implicit errors in agent traces, such as wrong clicks or citations when calls still succeed. The company says that on its benchmark flow-1 matches GPT-6-sol while costing 23 times less, so every production trace can be analyzed. flow-1 runs in a harness called Signals that treats spans like files, returns user-defined JSON findings, and clusters recurring events for eval datasets. Laminar’s observability stack is Apache 2.0 and self-hostable; revenue comes from managed cloud.

Unconfirmed: The quality and 23x cost claims are Laminar’s own benchmark results, not an independent evaluation.