Laminar's flow-1 checks every agent trace, not a sample, at 23x lower cost than GPT-6-sol

The newly released flow-1 moves agent-trace error checking from sampled inspection toward analyzing every trace, targeting implicit failures where each call succeeds but the agent clicks a wrong button or cites a wrong document mid-run. Prior practice either collected and stored traces, leaving data nobody had time to read, or used another LLM to inspect them; the latter is expensive at scale, and cheaper models often miss errors or flag unrelated things. On Laminar's own benchmark, flow-1 matches GPT-6-sol in detection quality while costing 23x less, the comparison that supports analyzing every production trace without sampling.

flow-1 runs inside a harness called Signals and treats an agent trace like a repository, so it can answer general prompts beyond raw error detection. The model was finetuned from a small open-source model and then post-trained with RL on thousands of synthetic investigations of realistic agent workflows in software, legal and customer support. It starts from a compressed string representation of the full trace plus a user prompt, greps across spans (an LLM call or tool call) and reads relevant ones in full, returning findings as JSON in a user-defined schema called signal events.

Once every trace is analyzed, the volume of individual events becomes its own burden, so the system automatically groups similar signal events into clusters. Reading individual events one by one is not feasible at production scale, and clustering compresses recurring problems into browsable categories. Users can drill from a cluster into individual events and their traces, build an eval dataset for that failure type from a cluster's traces, and use a coding agent to iterate on those evals before shipping a fix.

This report draws on the releaser's Hacker News launch post, covering flow-1's positioning, how the Signals harness works, and the clustering-to-eval loop. It is aimed at engineering teams running multi-step AI agents who need to inspect production traces. The post states no follow-up validation plan.

Both the detection-quality and cost comparisons come from Laminar's own benchmark, and the post reports no third-party reproduction. The post says cheaper models often miss errors or flag unrelated things, but gives no per-workflow breakdown of flow-1's performance.