# agent observability

Laminar released flow-1, a reinforcement-learning-trained model that finds errors in AI agent traces. Agents run many steps and can fail implicitly, clicking a wrong button or citing a wrong document while every call succeeds, and inspecting traces with a capable LLM is expensive at scale. On Laminar's own benchmark, flow-1 matches GPT-6-sol in detection quality at 23x lower cost, making it affordable to analyze every production trace without sampling.

Laminar released flow-1, a reinforcement-learning-trained model that finds errors in AI agent traces. Agents run many steps and can fail implicitly, clicking a wrong button or citing a wrong document while every call succeeds, and inspecting traces with a capable LLM is expensive at scale. On Laminar's own benchmark, flow-1 matches GPT-6-sol in detection quality at 23x lower cost, making it affordable to analyze every production trace without sampling.