Mistral expanded Studio on November 5, 2026, with managed agents for businesses automating work across multiple steps and connected systems. The Managed Agents service and its developer toolkit remain in public preview, with interfaces, behaviour and limits that may change.
Mistral describes agents that can keep working after an app closes, retain completed steps after an interruption and pause for approval. For an operations manager commissioning document processing or customer-service automation, the immediate decision is where the work runs and who holds its history.
Choose where the work and its records live
A worker is the software process that runs the agent and its tools. A session is the record of the job, including messages, tool calls and results. In the standard self-hosted deployment, these live in different places: your tools run on your infrastructure, but Mistral still stores the sessions.
The practical choices are:
Route | When it fits | What to confirm |
|---|---|---|
Hosted execution and Mistral-held job records are acceptable; you want Mistral to build and operate the worker. | Your team still defines the tools, access rules and approval process. Setup requires a connected GitHub repository and workspace access. | |
Tools need to run inside your network or under your own deployment process. | You operate the worker. Mistral still holds sessions and routes work; moving the worker does not make the whole system private. | |
The platform and its stored records—not just tool execution—must sit within a specified business environment. | Obtain an enterprise proposal covering the platform, model processing, supported features and operating responsibilities. |
Mistral says both the models and Studio can run in a customer’s cloud or on-premises environment. That is a separate deployment arrangement, not a consequence of starting a self-hosted worker. Confirm feature availability, processing locations, retention and support access in the actual proposal. RohitAI’s private-AI procurement guide offers a related framework for assigning those responsibilities.
For a short script or a single AI answer, a deployed agent may be unnecessary. The agent overview points single model requests to a simpler interface. Local agent development normally keeps the session in the running program; local execution does not by itself mean local model processing.
Include logs in the data decision
A trace is a step-by-step execution log. Mistral’s deployment documentation says tracing is enabled by default for both managed and self-hosted workers, including prompt and response content. Pattern-based redaction runs before that content leaves the worker, but can miss information in unfamiliar formats.
Teams can omit trace content or turn tracing off. Those controls are not described as turning off the separate session store. The practical implication: map the worker, model processing, saved sessions and traces separately before approving customer or employee information for a workflow.
Make human approval an explicit rule
Approval capability does not mean every consequential action waits for a person. Under the documented default permission mode, tool calls run if no permission policy is configured. A program can also answer an approval automatically. Specify which actions need a human, rather than treating the presence of an approval feature as that decision.
The timing matters. An unanswered approval waits one hour; the tool call then does not run, the model is informed and the current turn continues. This is different from a session’s lifetime: deployed sessions normally end after four idle hours or 24 hours total, whichever comes first. Waiting only for user input counts as idle.
For example, an approval raised after staff leave should not be assumed to remain pending the next morning. Session limits can be raised, but a session ended by a limit cannot resume. Match the separate approval and session clocks to the team’s working hours, and define what happens when nobody responds.
Plan for an action to be attempted again
Mistral’s recovery documentation says an interrupted deployed agent usually resumes after its last completed step. It can also fail or keep running until a limit. If a tool was interrupted with an unknown outcome, the model may call it again.
That distinction matters for a refund or customer notification: recovering the conversation is not a guarantee that the business action happens only once. Require repeat-safe actions—for example, a receiving system that recognises the same request identifier and returns the earlier result instead of issuing another refund.
Separate hosting access from the bill
For managed workflow hosting, a Pro or Team subscription alone is insufficient. The current deployment quickstart requires pay-as-you-go to be enabled for the organisation, or an Enterprise plan. Managed workflow hosting itself predates this release: Mistral’s release notes list its public preview on October 9.
The workflow deployment quotas allow three simultaneous managed deployments with pay-as-you-go enabled and ten for Enterprise. Free accounts and Pro or Team accounts without pay-as-you-go have zero. These are deployment slots—not simultaneous users, agent sessions or an included spending allowance. The published workflow quota should not be treated as a complete access table for every new agent feature.
Mistral’s billing guide says included monthly usage is shared across Studio, the API and Vibe Code. It is consumed first; further usage can draw on prepaid credits when pay-as-you-go is enabled. The organisation needs credits before enabling it.
The public pricing page does not establish an all-in price for a completed business process. Request a workload estimate identifying applicable model, tool, execution, sandbox, storage and support costs. A seat price or deployment slot cannot establish the cost of running the job.
Start with a process you can define
Choose one bounded task and name the person responsible for it, the actions that require approval, the permitted data locations and the response to a failed or repeated action. If the process needs fixed routing and mandatory steps, Mistral’s workflow layer is designed for explicit coordination; an agent can supply reasoning within that process.
Use managed execution when hosted processing is acceptable and operating a worker adds no value. Choose a self-hosted worker when tool placement is the requirement. If the requirement covers the platform, history and model processing too, take the private-deployment route to a concrete proposal before committing.