Launch HN: Laminar (YC S24) – We trained a model to catch silent agent failures — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Laminar released flow-1, a model fine-tuned and post-trained with reinforcement learning on thousands of synthetic agent investigations across software, legal, and customer support workflows to find errors in agent traces.
- flow-1 matches GPT-6-sol in detection quality while costing 23x less to run, which Laminar says makes it affordable to analyze every production agent trace instead of sampling.
- The model runs inside a harness called Signals that treats an agent trace like a code repository, with each span (LLM call, tool call) as a file — allowing general prompts such as checking whether a browser agent completed a checkout, not just raw error detection.
- Laminar automatically clusters similar signal events so recurring failures surface, and users can convert cluster traces into eval datasets to verify fixes before shipping them to production.
- Laminar is Apache 2.0 licensed and self-hostable at github.com/lmnr-ai/lmnr, with revenue from a managed cloud tier, and ships custom infrastructure including 20x trace compression and a custom SQL query engine.
Why it matters: Teams running AI agents in software, legal, or customer support often pay to store traces they never read. Laminar's flow-1 claims full-trace analysis costs less than sampling fewer traces with larger models — shifting observability economics for any shop that needs to catch silent failures where every API call still returns success.
Ask SkimNews
