Laminar releases flow-1, an RL-trained model that flags errors in agent traces

Laminar says flow-1 matches GPT-6-sol in detection quality on its benchmark while being 23x cheaper. The team fine-tuned a small open-source model, then post-trained it with RL on thousands of synthetic investigations of agent workflows in software, legal and customer support settings.

flow-1 runs inside a harness called Signals, which treats an agent trace like a code repository with each span, such as an LLM call or tool call, as a file. It starts from a compressed version of the trace and a user prompt, then greps across spans and reads the relevant ones in full. Signals can answer general prompts, such as detecting user frustration or checking whether a browser agent completed a checkout, and returns findings as JSON in a schema users define.

Laminar also groups similar findings into clusters that show how often each problem recurs. Users can turn the traces in a cluster into an eval dataset for that failure type and check a fix against those cases before release. Laminar is Apache 2.0 licensed and self-hostable, and the company makes money from its managed cloud.

Teams running agents in production can use this to review every trace for silent failures, such as a tool call that succeeds while the agent acts on bad data, instead of sampling traces or paying to store ones nobody reads.

The 23x cost figure and benchmark parity come from Laminar's own benchmark, so developers should test flagged findings on their own traces before relying on them.

Clustered findings can feed eval datasets, letting teams check that a fix works across real failure cases before shipping it.