Launch HN: Laminar (YC S24) – We trained a model to catch silent agent failures

Two founders launched an open-source observability stack for multi-step AI agents and released flow-1, a model trained with reinforcement learning to detect silent agent failures - errors that occur even when every call returns successfully (wrong clicks, incorrect citations, unfinished checkouts). On their benchmark flow-1 matches GPT-6-sol in detection quality while running 23× cheaper, making it affordable to analyze every production trace rather than sampling. The model was finetuned from a small OSS base and post-trained with RL on thousands of synthetic investigations across software, legal, and support workflows to specialize in trace analysis at low cost.

flow-1 runs inside a harness called Signals that treats an agent trace like a code repository: each span (LLM call, tool call, etc.) is a file, traces are compressed, and the model greps and reads relevant spans to answer general prompts. Outputs are structured JSON “signal events” in user-definable schemas, and similar events are automatically clustered so teams can find recurring failures, build eval datasets from clusters, iterate fixes with coding agents, and verify regressions before shipping. The stack includes 20× trace compression, a custom SQL query engine, automatic task extraction, is Apache 2.0 and self-hostable, and the company offers a managed cloud as its commercial product.