Mistral Studio managed agents: public preview, self-hostable

Mistral now runs your agents for you, and it will also run them inside your own walls. On November 5, Mistral expanded Studio with managed agents, a natural-language workflow builder and hosted apps. Per its release notes, Mistral Studio managed agents and the Agents SDK are in public preview. The pitch that sets it apart from other managed agent platforms: you can deploy the models and the platform in your own cloud, your own VPC or on-premises.

What actually happened

From Mistral's announcement and docs:

• Managed agents. You write an agent in Python with the Agents SDK. Studio runs it with skills, code sandboxes, subagents and hooks for custom logic at key points. You set limits on the tools, data and actions it can reach.
• Workflow builder. Describe a process in plain language and get a working workflow as editable code. Workflows mix fixed rules, agent reasoning and human approvals, and interrupted runs resume where they stopped.
• Tracing and evals. Managed agents send agent, model and tool spans to Studio by default. Offline and online evaluations run against your own datasets. Both are public preview.
• Connectors. Agents reach systems through connectors and MCP servers. Plain HTTP API connectors are in private preview.
• Deployment. Mistral-hosted, your cloud, your VPC or on-prem. Mistral names BNP Paribas as running Studio on its own infrastructure.

The agents docs say the API, behavior and limits "can change during the preview." Neither the announcement nor the docs publish a price for managed agents or sandboxes. Mistral's October notes say managed workflow deployments need a paid plan; we could not confirm the same for agents, so we do not assume it.

Why it matters for your business

Most managed agent platforms make you send customer data to the vendor's cloud. For a clinic, a law office or a distributor with a strict customer contract, that is the deal-breaker. A platform you can run in your own VPC removes that objection.

The catch is the word "preview." Behavior and limits can change under you, and there is no price sheet yet. That is fine for an internal pilot. It is not fine for the order-routing agent your warehouse depends on.

If you test it: pin the SDK version, keep your prompts and tool definitions in your own repo, and write the workflow so the model call is one swappable step. Then a pricing surprise in GA is an afternoon of work, not a rebuild.

Sources: Mistral AI, Mistral release notes, Mistral docs.