Mistral Expands Studio With Managed Agents, Workflows and Apps: Its Docs Say Managed Agents and the Agents SDK Are a Public Preview, Sessions End After 24 Hours by Default, Self-Hosted Mode Still Has Mistral Store the Session, and Time Savings Are Self-Reported

On November 5, 2026, Mistral AI said it is expanding Mistral Studio, its enterprise platform, with managed agents and apps, plus new features for workflows, enterprise context, agent observability and evaluations. The post is framed around the Mistral Large 4 model, which it nicknames Le Chonk, and says customers such as ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services already use Studio in production. One fact is missing from the post and appears only in the documentation: the Managed Agents page says Managed Agents and the Agents SDK are in public preview, and that their API, behavior and limits can change and may require code changes.

Managed agents are the core of the release. Mistral says developers can build agents for multi-step tasks, give them skills that capture a team's procedures, sandboxes where they can run code, and subagents for parts of a task, and add hooks for custom logic at key points. Studio runs the agent on infrastructure that Mistral says supports millions of Mistral Vibe sessions each week, a usage figure it does not break down. The documentation describes an agent as a model plus a harness, with a session that records each message, tool call and result and an optional environment, which is a sandbox. By default an agent has no environment.

The documentation also gives the operating limits the post leaves out. A deployed task can run for days, but by default a session ends after 24 hours, or after 4 hours idle when no turn is running or the turn is only waiting for a person, and both limits can be raised. If the machine fails, the turn resumes from its last finished model call or tool call. A turn that waits for an approval waits one hour by default, even when the app is closed. Each turn is a trace in Studio with one span per model call and tool call. There are two ways to deploy: `agents serve` runs the harness and the tools on your own servers, and `agents deploy` has Mistral run them.

Workflows and apps are the other two pieces. A workflow defines which steps follow fixed rules, where an agent reasons and where a person decides, and an interrupted run can resume. Mistral adds event triggers from enterprise systems and a visual graph, and says a workflow can be generated from a description in Mistral Vibe, as code that teams can inspect and edit. Apps get hosting, authentication and access controls, built from pre-approved reusable components so less new code needs security review. A new Mistral MCP server lets a developer build with Mistral Vibe or another coding agent, test locally and deploy to Studio with one command.

The sovereignty claim needs care. The post says the models and the platform can run in your own cloud, in your VPC or on premises, or in Mistral-hosted deployments, and that sensitive data can stay inside your perimeter. The documentation is more specific about the managed-agent mode: with `agents serve`, the harness and your tools run on your servers, but Mistral stores the session. The sandbox can also be hosted by Mistral. The model can run on Mistral's API or behind your own gateway. Which setup keeps which data where is therefore a question to put to Mistral before a regulated team relies on the word sovereign. We did not test any deployment.

The savings figures are Mistral's own and have no stated baseline. An incident-investigation agent built by its engineers saves the team "hundreds of hours each month", and a feedback-routing workflow saves its product operations team about 35 hours a week. CMA CGM says each new process is production-ready in weeks rather than months. These are vendor and customer self-reports; the reviewed post gives no method, baseline or reproducible data, so we cannot independently verify them. The post gives no benchmark, no latency or cost figure, no price, and no list of regions or plans.

Our reading: the release is a real product step toward agents that survive crashes, wait for approvals and leave a trace, and the documentation makes it more checkable than the post alone. Teams that want to try it should read the preview notice first, plan for API changes, set the session and approval limits on purpose, and decide which mode stores what, before moving a sensitive process onto it.