Mistral AI expanded its enterprise platform Mistral Studio on November 5, 2026 with managed agents and apps, both in public preview, plus new workflow, observability and evaluation features.
Managed Agents are built in Python with the Agents SDK. The same code can run three ways: inside the developer's own process, self-hosted with the agents serve command, or Mistral-managed with agents deploy.
Once deployed, execution is durable. Mistral says an agent can work for hours or days, resume from its last finished step after a crash and continue after the calling app exits. A turn can wait for a person's approval and resume later. Developers choose which tool calls need approval.
Agents get skills, per-session sandboxes for running code, subagents, connectors, MCP servers and custom hooks. Each turn appears in Studio as a trace with one span per model call and tool call.
Apps adds hosting, authentication and access controls for internal applications built on agents and workflows. A workflow builder in public preview turns a plain-language description into workflow code that teams can inspect and edit. HTTP API connectors, in private preview, let agents call web APIs without an MCP server.
Mistral says Studio and its models can run in Mistral's cloud, a customer's VPC, on premises or a mix. It names ABANCA, BNP Paribas, CMA CGM and Tata Consultancy Services among Studio users. BNP Paribas says it runs Studio on its own infrastructure.
Why this matters
Long-running agents fail in the middle, not at the start. Durable state, resumable steps and approval gates are what move an agent from a demo to an operated system. The on-premises option keeps that control inside the customer's own perimeter.
Sources: Mistral announcement, Mistral release notes.
This article is not legal advice.
