Mistral Studio adds managed agents, hosted apps and new workflow tools for enterprises

Mistral AI on November 5 broadened Mistral Studio, its platform for building and running AI systems in production, with managed agents and apps alongside new workflow, observability and evaluation features. The company presents the update as the way to put Mistral Large 4, its open-weight flagship nicknamed "Le Chonk", to work on business tasks.

What Studio now includes

Managed agents are designed for multi-step jobs. Developers can equip them with skills that encode a team's procedures, sandboxes for running code, subagents that take over parts of a task and hooks that insert custom logic at set points. Mistral operates the runtime, while customers set limits on which tools, data and actions an agent may use.

Workflows let teams decide which steps follow fixed rules, where an agent reasons and where a person makes the call; interrupted runs can resume. New additions include triggers fired by events in enterprise systems and a visual graph of the logic.

Apps are the third piece. Studio supplies hosting, authentication and access controls, and apps can call agents and workflows, query search indexes and read shared enterprise context. Mistral says they deploy with a single command, either to a dedicated link or inside Mistral Vibe. Runs are traced step by step with OpenTelemetry, and evaluation tools let teams test prompt, model or tool changes on their own datasets before release.

Deployment and availability

Customers can run the stack on Mistral's infrastructure, in a cloud, in their own VPC, on-premises or in a mix. The announcement links to a pricing page but gives no figures. According to AI/TLDR, most of the new features are in public preview, meaning APIs and limits may still change, while Studio apps inside Vibe Code Web are generally available.

Mistral lists ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services as customers or partners. Its own examples claim savings of hundreds of hours a month on incident investigations and about 35 hours a week on triaging internal feedback; these are company estimates without independent verification.

Why it matters

RuntimeWire reads the move as an attempt to own the layer that connects models to corporate data, permissions, approvals and audits, competing with OpenAI, AWS and Cohere for regulated clients. How the platform performs at scale, and what it will cost once out of preview, are still open questions.