Mistral will now run your company's agents, or hand you the whole stack to run yourself

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 5 November Mistral expanded Mistral Studio, its enterprise AI platform, with managed agents and apps, alongside new workflow, enterprise context, observability and evaluation features. It positions the platform around its Mistral Large 4 model, which it nicknames Le Chonk, and describes the split simply: "An agent investigates and acts, a workflow coordinates execution and approvals, and an app gives employees a way to use these systems."

The release notes say managed agents can run for hours or days, resume from their last finished step after a crash, and pause a turn to wait for a person's approval. A workflow builder turns a plain-language description into code that can be inspected and edited, and new HTTP connectors let agents call web APIs without anyone building an MCP server first. Mistral says customers choose which actions require human approval, OpenTelemetry traces show every model call, tool use and approval, and evaluations test changes against the customer's own datasets. Models and platform can be deployed in the customer's own cloud, VPC or on-premises. Mistral's own examples include an incident-investigation agent that asks for approval before sensitive actions, and it says customers use Studio for work from penetration testing to KYC onboarding. BNP Paribas's head of AI and IT innovation is quoted saying the bank runs Studio "on our own infrastructure".

Why it matters

Every serious agent platform now promises approvals, traces and evaluations. The difference here is where they run. For banks that cannot send customer data to a foreign cloud, or that want to freeze a model version for validation, a full agent stack that can sit inside their own perimeter is a real option, not a compromise.

Approvals, traces and evaluations are only controls if they stay in your hands after the vendor leaves the room.

Two cautions. First, much of this is in preview: managed agents, apps and the workflow builder are public previews, and HTTP connectors are a private preview. Firms should not build regulated processes on preview features without a fallback. Second, "you decide which actions require human approval" puts the design burden on the customer. In my view the approval list deserves the same scrutiny as a delegation-of-authority matrix: which actions, above which values, approved by whom, and recorded where. The traces will tell you what happened. They will not tell you whether the approval rules were right.

What I'd watch

• When managed agents and apps reach general availability on private-cloud and on-premises deployments.
• Independent evaluations of Mistral Large 4 on regulated-document and KYC tasks.
• Whether European banks publish production results from Studio-built agents, beyond quotes.