Laminar launches flow-1, open RL model that triages agent traces

On its own benchmark, Laminar's new flow-1 model processes 888 agent traces per dollar, 23 times what GPT-6-sol can do at the same price, while matching that model's quality at catching silent errors, the YC S24 team says.

Co-founders Robert and Din posted the launch on Hacker News. 'Agents now run for many steps and fail implicitly,' they wrote. 'An agent can click the wrong button or cite the wrong document while every call completes successfully. These errors are the most useful ones to catch and understand.' flow-1 is pitched as a triage layer that reads traces end-to-end rather than sampling them.

The model was finetuned from an unnamed open-source base and 'post trained it with RL on thousands of synthetic investigations of realistic agent workflows in software, legal, customer support, etc.' The founders wrapped it in a harness they call Signals, which treats a trace 'like a repo, with each span (an LLM call, tool call, etc.) as a file,' using grep and read against the spans and returning JSON signal events. On the company's benchmark page, Laminar says flow-1 'surpasses GPT-6-sol with high reasoning in failure detection quality' at $0.0011 per trace versus $0.026.

The platform ships under Apache 2.0 from lmnr-ai/lmnr; Laminar monetizes a managed cloud. The benchmark numbers come from the vendor, not from an independent evaluation. It lands in a crowded beat: our agents tracker has logged 500 stories in 90 days.