Robert and Din from Laminar, a Y Combinator S24 company, have released flow-1, an open-source observability model trained with reinforcement learning to find errors in AI agent traces. On the company's benchmark, flow-1 matches GPT-6-sol in detection quality while being 23x cheaper, making it affordable to analyze every production agent trace rather than sampling.
The team argues that agents now run for many steps and often fail silently. An agent can click the wrong button or cite the wrong document while every call completes successfully. Catching these errors requires understanding what the agent was trying to do and inspecting its LLM calls and tool results, not just reading the final output. Collecting traces alone does not solve the problem, since many platforms are essentially data collectors that store traces nobody reads.
flow-1 runs inside a harness called Signals, which acts like a coding agent that treats an agent trace as a repository, with each span treated as a file. It greps across spans, reads relevant ones in full, and returns findings as JSON in user-defined schemas. Similar signal events are grouped into clusters, which users can use to build evaluation datasets for fixing recurring failures.
Laminar is licensed under Apache 2.0 and can be self-hosted, and the company earns revenue from its managed cloud. The founders asked the community how they track agent performance and how they verify that fixes do not introduce new failures.
Source: Hacker News · Summarized by HeadlinesBriefing