Laminar launches flow-1 model to catch silent AI agent failures

The company said agents now run for many steps and fail implicitly: an agent can click the wrong button or cite the wrong document while every call completes successfully. Catching such errors requires understanding what the agent was trying to do and inspecting its LLM calls and tool results, Laminar said, because looking only at final output is insufficient. Collecting traces alone does not solve the problem, according to the post; many agent observability platforms remain data collectors at their core, and customers can end up paying to store traces nobody has time to read. Using another LLM to inspect traces is the solution, Laminar said, but capable models are expensive at scale and cheaper ones often miss errors or flag unrelated things.

To address that, Laminar finetuned a small open-source model and post-trained it with RL on thousands of synthetic investigations of realistic agent workflows in software, legal, customer support and other areas. The resulting model runs inside a harness called Signals. Signals acts as a coding agent, treating an agent trace like a repository, with each span—an LLM call, tool call or other step—as a file. It starts from a compressed string representation of a full trace and a user prompt, then greps across spans and reads relevant ones in full.

Laminar said Signals can respond to general prompts, such as identifying user frustration or checking whether a browser agent completed a checkout, and is not limited to raw error detection. It returns findings as JSON in a schema users define; Laminar calls these signal events. Because analyzing every trace produces too many events to read, Laminar automatically groups similar signal events into clusters, showing recurring problems and their frequency and allowing users to drill into individual events and the traces behind them.

Traces from a cluster can be used to build an evaluation dataset for that type of failure. Teams can then use a coding agent to hillclimb on those evaluations and check that a fix works across the cases before shipping to production. Laminar describes the intended loop as trace, analyze, catch errors and form clusters, fix, then release. The company said making that loop easy is why it built Laminar.

Laminar said it has also built 20x trace compression, a custom SQL query engine and automatic task extraction, and that its blog explains how close inspection of agent behavior led to those components. The core product is Apache 2.0 and self-hostable, with the company making money from a managed cloud service. In the Launch HN post, the founders asked how teams track agent performance today, how much they learn from user feedback versus their own monitoring, and how they check that a fix has not introduced a different failure.