Mistral Studio Adds Agents That Run for Days

Mistral AI on Thursday, Nov. 5, expanded its Studio platform with managed agents that can run for hours or days, a natural-language workflow builder, a command-line tool for internal apps and a set of evaluation tools. The French company put most of the new features into public preview, pairing them with its open-weight Mistral Large 4 model to compete for enterprise agent workloads.

Key takeaways

• Managed agents built with Mistral's new Python Agents SDK can run in a developer's own process, on self-hosted servers or on Mistral's infrastructure, and resume from the last finished step after a crash.
• A workflow builder lets users describe a process in plain language; Mistral's Vibe coding agent then writes, tests and deploys the code, which stays editable.
• Pricing, regions and general-availability dates for the new features were not disclosed; nearly all are labeled public preview.

How Mistral Studio managed agents work

According to Mistral's Nov. 5 release notes, managed agents follow a "write an agent once, run it anywhere" model. Developers write the agent in Python using the Agents SDK, published as the mistralai-agents package, then choose where it runs.

An agent can run inside the developer's own process, on their own servers with the "agents serve" command, or on Mistral's infrastructure with "agents deploy." Deployed agents run in the background on durable execution, so a job that "works for hours or days" picks up from its last completed step if something crashes. Mistral hosts a separate sandbox for each session.

The company says the runtime rests on the same infrastructure that supports "millions of Mistral Vibe sessions each week," referring to its coding agent.

Workflow builder, Apps CLI and evaluations

The new workflow builder takes a plain-language description of a business process. In Mistral's words, "Vibe writes it, tests it, and deploys it," and users can inspect and edit the generated code. A read-only graph view of workflows arrived earlier, with a Workflow Visualizer preview on Oct. 9.

For internal tools, the Mistral Apps CLI adds commands to scaffold an app ("mistral apps init"), run it locally ("mistral apps dev") and publish it into Studio ("mistral apps deploy"). On the testing side, an Evaluation SDK supports offline evaluations, managed datasets uploaded as CSV or JSONL files, datasets built from production telemetry and online evaluations on live traffic.

Other Nov. 5 additions include OpenTelemetry-based tracing that shows model calls, tool use and approvals, a managed search index, and webhooks and triggers. HTTP API connectors are in private preview, while Studio apps inside Vibe Code Web are generally available.

Who is using it

In its announcement post, titled "Le Chonk brings the brains. Studio puts it to work.," Mistral named ABANCA, Accenture, Alten, BNP Paribas, Capgemini, shipping group CMA CGM and Tata Consultancy Services as companies running Studio in production. Customers can deploy both the models and the platform in their own cloud, a virtual private cloud or on premises, use Mistral-hosted deployments, or combine the two.

Mistral also cited internal examples: an incident-investigation agent it says saves its engineering team "hundreds of hours each month," and a workflow for sorting submissions that saves its product operations team about 35 hours a week. Julien Smadja, vice president of AI at CMA CGM, said in the post that "each new process we automate is production-ready in weeks rather than months."

Context: Le Chonk meets the agent platform race

"Le Chonk" is the nickname for Mistral Large 4, the roughly 1-trillion-parameter model whose weights Mistral published after raising €3 billion at a €21 billion valuation. Studio is the commercial layer meant to turn that model into revenue, and its pitch leans on control: companies in regulated sectors, such as banks, can keep both the model and the agent runtime inside their own environment.

What's next

Mistral has not said when the preview features will become generally available or how they will be priced. Until then, enterprise buyers will be judging whether the durable runtime and built-in evaluations are mature enough for what BNP Paribas AI chief Su Yang described in the post as "some of our most sensitive processes."

FAQ

What are Mistral Studio managed agents?
They are agents written with Mistral's Python Agents SDK that can run locally, on self-hosted servers or on Mistral's infrastructure, with durable execution that resumes after a crash.

Can Mistral Studio run on premises?
Yes. Mistral says customers can deploy both its models and the Studio platform in their own cloud, a VPC or on-premises environment, or use Mistral-hosted deployments.

How much does it cost?
Mistral did not publish pricing for the new features in its Nov. 5 announcement or release notes.

References

• Le Chonk brings the brains. Studio puts it to work. — Mistral AI (Nov. 5, 2026)
• Release notes — Mistral Docs (Nov. 5, 2026)