Launch HN: Laminar's flow-1 matches GPT-6-sol at catching silent agent failures, 23x cheaper

skull8888888 · hn · 2026-10-20

Laminar (YC S24) released flow-1, an RL-trained model built to catch "silent failures" in AI agent traces — cases where every call succeeds but the agent clicked the wrong button or cited the wrong document.
The team finetuned a small open-source model, then post-trained it with RL on thousands of synthetic investigations of realistic agent workflows across software, legal, and customer support.
On their benchmark, flow-1 matches GPT-6-sol in detection quality while being 23x cheaper, making it affordable to analyze every production agent trace instead of sampling.
Their thesis: collecting traces alone is worthless since most observability platforms are data collectors; using an LLM to inspect traces is the answer, but capable models are too expensive at scale and cheap ones miss errors. flow-1 runs inside a harness called Signals.
The project is open source, with a blog of example trace analyses and a one-minute walkthrough video.

More from coding & agent