Conntrail is an observability tool for agentic workflows. It traces the decisions an agent makes and measures how stable those decisions actually are, using contrastive perturbation entropy (CPE) to show when a routing choice was confident versus when it was one paraphrase away from going a different direction.

Most observability tools tell you what an agent did. Conntrail also tells you how fragile that decision was, which part of the input drove it, and what the agent would likely have done instead if that part changed.

Conntrail is the standalone, self hostable version of the tool: an ingestion service, a persistent trace store, a query API, and a dashboard for exploring traces, entropy scores, and failure classifications across any agentic workflow you point it at.

Conntrail Lib is the original prototype library. Its tracing core, contrast generation, divergence analysis, and entropy scoring, is real and working, and is the foundation the standalone tool is built from.

Conntrail is under active development. The core tracing math is proven; the ingestion service, trace store, dashboard, and deployment story are being built out now.

Contributions are welcome. If you are interested in agent observability, routing stability, or want to help build out the collector, dashboard, or SDK integrations, feel free to open an issue or a pull request on either repository.