ZenML rebrands as ZenML Labs and adds AI engineers to its platform

One company, three products

ZenML Labs said the rebrand reflects its focus on helping teams run AI systems in production. ZenML remains the company’s AI orchestration product, while Kitaru is designed to identify gaps in agent evaluations by examining production sessions. ZenML Pro is the managed platform spanning both products, with AI engineers built on top.

ZenML said teams including JetBrains, Safran and Adeo have used its technology to run AI in production for five years. The company now sees agents creating similar operational challenges, including silent failures, unexpected customer requests, prompt changes that affect previous behavior and rising model costs.

AI engineers inspect traces and test fixes

The new agent harness in ZenML Pro comes with ZenML and Kitaru skills designed for work such as writing evaluations, investigating failures and testing changes. The company describes each bot as an AI engineer for a particular agent, model or job. These bots provide evidence of their work, while people decide what is released.

According to ZenML, the system reads production traces alongside existing evaluations, groups similar failures and highlights the sessions most worth reviewing. It can also connect to GitHub to locate an agent, its evaluations and relevant policies. ZenML said users can revoke the keys used to access repositories and traces, and that it stores only metadata. Teams that need data to remain in their own environment can run the agents in their own cloud with their own large language model endpoint.

When a possible fix is found, the agents can run it as an experiment against past production sessions. ZenML’s example describes a claims agent where 31 of 214 appeals ended without a decision after a switch to a cheaper model, even though an evaluation checking response format passed every case. The company says its AI engineers could identify the pattern and draft a decision_made check.

Roadmap includes model fine-tuning

ZenML is positioning production traces as potential training data. On its roadmap, the company says its AI engineers will fine-tune a cheaper open model on those traces through a ZenML pipeline running on the customer’s infrastructure. The resulting model would be evaluated against the same checks, with the customer approving any switch and receiving the weights.

For existing ZenML users, the company says pipelines will continue running with the same SDK, stacks and integrations, without code changes. The dashboard has a new design, while the classic dashboard remains available at classic.cloud.zenml.io. ZenML and Kitaru remain open source under Apache 2.0, and the company says self-hosted and on-premises deployments will move to the new dashboard when upgraded.

ZenML Pro users received the new interface the week before the announcement, and Kitaru’s new interface ships today, according to ZenML. The company’s direction is to let people define what better means while AI engineers investigate failures, test improvements and eventually train and deploy models.