Mistral Studio now runs your agents for you, on its cloud or on yours. The hosting choice is the real decision

Most agent platforms ask you to choose a vendor and accept its cloud. Mistral’s latest release asks a different question: once you have chosen the platform, where do you want it to run?

What Mistral added to Studio

Mistral Studio, the company’s platform for building enterprise agents, workflows and apps, now includes a managed runtime. In the November release:

• Managed agents can be given skills, code sandboxes, subagents and hooks, while Studio runs them.
• Workflows can resume after an interruption, start from events in enterprise systems, and are shown as a visual graph.
• Apps get hosting, authentication and access controls, and can be assembled from pre-approved components.
• Observability and evaluation come as OpenTelemetry traces and tests against the customer’s own datasets before release.
• A Mistral MCP server lets developers build with Mistral Vibe or an existing coding agent, test locally and deploy with one command.

Mistral says the same runtime supports millions of Mistral Vibe sessions a week, and cites ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services as production users of Studio. The productivity examples it gives, such as an incident-investigation agent saving hundreds of engineering hours a month, come from Mistral’s own internal teams.

One platform, four operating models

The part that sets Mistral apart is the deployment list: Mistral-hosted, the customer’s own cloud, a VPC, on-premises, or a combination.

Where Mistral Studio can run, and who carries the work

Mistral operates itYou operate it

1. Mistral-hostedFastest start; data and runtime on Mistral's side
2. Your cloud accountYour contracts and region; shared operations
3. Your VPCNetwork isolation; your team handles more of the stack
4. On-premisesFull custody; upgrades, capacity and uptime are yours

This matters because the hard parts of an agent platform are operational. Durable workflows hold state across failures. Sandboxes execute code. Traces contain prompts and tool results, which often include customer data. Whoever runs the runtime holds all of that.

For a European bank, that is the whole question. A platform that can run on-premises lets a supervised institution keep agent state, traces and evaluation data inside its own perimeter, which is easier to defend to a supervisor than a contract clause about a vendor’s cloud. It also means the bank now runs a complex, fast-moving platform with its own staff, at the vendor’s release pace.

We saw the same trade-off with Mistral Large 4 and its move to open weights: openness gives you the option to run it, and the duty that comes with it.

How to choose

Three questions settle most cases.

What does the trace contain? If agent traces will hold regulated personal data or trade secrets, the trace store’s location decides the deployment, not the model’s.

Who will patch the runtime? A managed runtime with sandboxes and workflow state needs security updates on the vendor’s schedule. If no team can commit to applying them within days, self-hosting creates more risk than it removes.

Which agents need which option? The combination option is the realistic one. An internal IT helper can run hosted; an agent that reads customer accounts can run in the VPC. Decide per agent class, and keep the evaluation datasets and traces in the most restrictive location the agents share.

The release confirms a pattern we described in our guide to the production enterprise AI stack: the agent runtime has become infrastructure. Mistral now sells it in four sizes of ownership. Pick the one your team can actually operate.