ZenML

ZenML builds software for teams running machine learning pipelines and AI agents in production. Its ZenML framework orchestrates training, batch inference and evaluation, with versioned artifacts and a model registry. Kitaru supports agent evaluations: teams import production traces, organize cases around failure modes, incorporate domain experts’ judgments and test changes against recorded sessions. Its replays execute the actual agent code, with a history policy that substitutes recorded tool responses for live calls.

Co-founders Adam Probst and Hamza Tahir connect the company’s ML infrastructure work with its agent tooling. ZenML’s technical approach separates a metadata control plane from customers’ data, artifacts and compute, which remain in their infrastructure. Pipeline code can run locally and on Kubernetes unchanged, while artifact versioning and lineage track outputs across steps and runs. Kitaru extends this emphasis on reproducibility to agent behavior, keeping evaluators in code and comparing agent versions against versioned groups of test sessions.

The business combines free, self-hosted open-source software with paid managed control planes and enterprise governance features, including access controls and audit logs. A concrete deployment is JetBrains: ZenML reported that its customer’s usage grew from eight active users in April 2025 to 44 active users in October, when it ran 49 unique pipelines and 2,765 monthly runs.