Agent observability has a labeling problem. When an agent clicks the wrong button or cites the wrong document, every API call returns 200. The trace looks clean. The error is semantic, not syntactic. You need a human to read the entire execution log and decide whether the agent actually did what it was supposed to do.

Laminar built flow-1, a small open-source model trained with reinforcement learning to find these implicit failures in agent traces. It runs as a coding agent inside a harness called Signals, treating each span (LLM call, tool invocation, retrieval step) as a file in a repo. The model starts from a compressed string representation of the full trace and investigates like a developer debugging a failed CI run.

The interesting part is not that they use an LLM to analyze traces. The interesting part is how they trained a cheap model to do it without ground-truth labels.

The Implicit Failure Problem

Traditional observability assumes failures are explicit. An HTTP 500, a timeout, an exception.…