Mistral promises AI agents on your own servers, but not before late November

At Mistral AI, an AI agent already handles investigations into production outages: it follows the team's procedure guide and asks for approval before any sensitive action. The French start-up chose this example on November 5 to present the new version of Studio, its platform for businesses. It says the agent saves "hundreds of hours" a month, without detailing how it calculated that.

Studio can now be used to build, deploy and monitor agents (programs that chain several tasks on their own), automated workflows and internal applications, based on Mistral Large 4. The central pitch: everything could run in the cloud, in a private cloud or on the customer's own servers.

What the platform adds

Agents get skills that describe the team's procedures, isolated sandboxes to run code, and sub-agents. Workflows can be triggered by the company's software and resume after an interruption. Every step is logged, human approvals included, and changes can be tested on the customer's data. Mistral cites seven customers in production, including BNP Paribas, whose head of AI says the bank runs Studio "on our own infrastructure." These testimonials are published by Mistral and have not been verified.

On your servers, but not right away

The November 5 release notes label almost all the new features as public preview, hosted by Mistral. According to a Mistral message relayed by AlphaSignal, the private cloud and on-premises versions will arrive in late November, with no precise date. Until then, the documentation already lets you run an agent on your own machines, but orchestrated by Mistral. An October 9 release note states that, during the preview, managed workflow deployments require a paid plan and are subject to a quota. No price has been announced. The same day, Mistral acquired SkipLabs, a specialist in the reliability of AI-generated code, its fifth acquisition of the year according to Maddyness.

Before handing the keys to an agent

One detail in the documentation deserves attention: by default, an agent runs its tool calls without asking for anything as long as no authorization policy is defined. Before connecting an agent to a real system, it is therefore best to:

• switch to approval mode for sensitive actions;
• limit the tools and data it can access;
• reserve the preview for testing, as the interface may still change.

We'll check back in late November to see whether the versions installed at customers' sites arrive on time, and at what price.