Mistral Studio adds managed agents and apps for enterprise AI teams

Mistral has added a managed agent runtime, hosted internal apps, a workflow builder, tracing and evaluations to Studio, its enterprise AI platform, in a release dated November 5, 2026. The most important fact for builders: almost every new piece is a public preview hosted by Mistral, with no published price for hosted runs, and Mistral warns that the preview API can still change.

The launch extends Workflows, the orchestration layer Mistral first put into public preview in April, into a broader platform for agents that run for hours or days, pause for human approval and resume after crashes. Coverage disagrees on one headline point, whether Workflows itself is now generally available (GA, meaning out of preview), and on how far the platform can run inside a customer’s own infrastructure.

What Mistral added to Studio on November 5

According to coverage of Mistral’s announcement, Studio now runs agents that come with skills (written team procedures), code sandboxes, subagents for delegating tasks and hooks for custom logic, with Studio operating the runtime. Workflows mix fixed rules, agent reasoning and human approvals, can resume after an interruption, and can start from events in enterprise systems. Apps get hosting, authentication and access controls and deploy with one command.

Observability and testing are part of the release. Runs emit OpenTelemetry traces (an open standard for recording what software did) covering model calls, tool use and approvals, and teams can test a prompt, model or tool change against their own datasets before release. Evaluations come in two forms: offline evaluations from the mistralai-evaluations SDK and online evaluations on live traffic, with datasets built from CSV, JSONL or production traces.

Two developer-tool additions sit around the platform. Builders can generate a workflow from a plain-language description inside Mistral Vibe, and a new Mistral MCP server lets other coding agents build and deploy to Studio. MCP, the Model Context Protocol, is a standard way for AI tools to connect to external services. The workflow builder in Vibe writes, tests and deploys a workflow, with editable code and a live workflow graph.

The announcement leads with Mistral Large 4, though customers choose their own models.

How Managed Agents work

Managed Agents are Python agents built with the Agents SDK, published as mistralai-agents under the Apache-2.0 licence. A single agents.Agent(...) definition holds the model, instructions, typed tools, connectors, MCP servers, skills, secrets, guardrails, hooks and structured output.

One codebase supports three run modes:

• In-process, for development.
• agents serve, where “the harness and your tools run on your servers” while Mistral stores the session so it survives a crash.
• agents deploy, where Mistral hosts the agent, building a pushed GitHub commit from the project’s Dockerfile.

Deployed sessions can run for days and resume from the last finished model or tool call after a worker crash; because Mistral stores them, teams do not host a database. Approvals can be set per tool call and answered by code or left for a person, with the turn waiting even while the app is closed, one hour by default. Running turns can be steered or interrupted, and several clients can connect to one session for live views and approvals.

Subagents run side by side on parts of a task, with one approval policy covering the whole tree and limits on concurrency and depth. Sandboxes are either a Mistral-hosted Linux container per session with Python, a shell and a writable workspace, or local Docker and in-process sandboxes for development and CI, with outbound network policies. Skills are SKILL.md bundles loaded on demand, and Mistral connectors act with the session user’s own permissions.

Workflows: from April preview to a contested GA label

Mistral launched Workflows in public preview in Studio on April 28, describing it as an orchestration layer for running AI-powered processes in production: connecting tools, managing multi-step pipelines, handling failures and pausing for human approval without losing progress. A developer writes the workflow in Python and publishes it to Le Chat, Mistral’s assistant, so anyone in the organisation can trigger it, and every step is tracked and auditable in Studio. Workflows is built on Temporal’s durable execution engine.

Mistral’s head of product, Elisa Salamanca, described the design as hybrid at the April launch:

“So Workflows has been built in a way that you can orchestrate things deterministically and inject agentic pieces whenever you want them to be done, so it combines deterministic code with agentic capabilities.”

She told The Deep View that this reflected a focus on “harder problems that need orchestration happening in different places where the execution is happening”. Use cases cited then included cargo release automation, document compliance checking and customer support triage.

On the current status, the sources do not line up. One review notes that launch coverage described Workflows as generally available, while Mistral’s Workflows documentation still shows a public preview notice, and advises checking the status for each account. The release-notes table reproduced by beri.net lists the workflow builder as public preview. Builders should treat the GA claim as unconfirmed until it appears on their own account and in Mistral’s docs.

What is preview and what is GA

Mistral’s release notes list 12 entries dated November 5, 2026, and every new piece is labelled Mistral-hosted, according to beri.net, which reproduces the following:

Feature	Status	Hosting
Managed Agents (SDK)	Public preview	Mistral-hosted
Workflow builder	Public preview	Mistral-hosted
HTTP API connectors	Private preview	Mistral-hosted
Mistral Apps CLI	Public preview	Mistral-hosted
Offline evaluations, managed datasets, datasets from telemetry	Public preview	Mistral-hosted
Managed search index	Public preview	Mistral-hosted
Tracing for managed agents	Public preview	Mistral-hosted
Webhooks and triggers	Public preview	Mistral-hosted
Online evaluations	Public preview	Mistral-hosted
Studio apps in Vibe Code Web	GA	Mistral-hosted

The Managed Agents docs warn that the “API, behavior, and limits can change during the preview,” so code written now may need changes before GA. HTTP API connectors, which would let teams register any web API without first building an MCP server, are private preview and invitation only.

Access, plans and pricing gaps

Mistral-hosted managed deployments require pay-as-you-go or an Enterprise plan. Mistral’s quota page lists 0 concurrent managed deployments on Free and on Pro or Team without pay-as-you-go, 3 with pay-as-you-go, and 10 on Enterprise. The Agents SDK is Python-only, requiring Python 3.12 or later, and hosted deployment builds from a GitHub repository through the GitHub connector.

Mistral’s public pricing pages list per-token model prices but no itemised prices for Mistral-hosted sandboxes or managed deployments. beri.net describes the release as a public preview “with no published price”.

Mistral names seven production users: ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and TCS. In the launch post, BNP Paribas says “we run Mistral Studio on our own infrastructure and keep those systems under our control,” and ABANCA calls Studio “our platform for building enterprise AI applications”.

Where the data goes: self-hosted is not on-premises

For teams choosing Mistral partly for data control, the hosting split matters most. When a workload is self-hosted, Mistral still handles routing work to the right deployment, durability across crashes and restarts, load-balanced scaling, and the executions, logs and traces shown in Studio. Mistral’s self-hosted deployments page states: “Self-hosting a deployment is not the same as running the whole Studio platform on your own premises”.

Agents and workflows can be self-hosted in this way; apps cannot, because they “run only on Mistral Cloud”. For workflows, customer code runs on workers that connect outbound only, while Mistral runs the orchestrator that holds state, history and task dispatch. Self-hosted agent workers also rely on Mistral’s API for model calls, so they are not an air-gapped option.

Two details reduce the exposure. The SDK encrypts payloads before they leave the worker, so Mistral stores ciphertext, and inputs or outputs over 2MB go to the customer’s own S3, GCS or Azure blob storage with only a reference kept. Tracing is on by default, with redaction applied before export, covering common secrets and personal data. beri.net advises checking what that redaction removes before pointing an agent at customer records.

Running the whole platform in a private cloud or on premises is for enterprise customers, and launch coverage reports private-cloud and on-premises availability. beri.net describes the fully on-premises Studio as a separate Enterprise contract. The sources reviewed do not show a confirmed date for broader private-cloud or on-premises rollout.

What it means for builders

For teams already on Mistral’s older agent tooling, there is migration work ahead. Mistral’s docs now label the previous Agents API overview as deprecated, “preserved for existing integrations,” and direct new builds to the Agents SDK. That older API offered persistent conversation state, built-in connectors for code execution, web search, image generation and document libraries, and handoffs between agents. DataStudios reports that the Agents and Conversations API (client.beta.agents) sits under a Deprecated section with no removal date, according to beri.net.

The new SDK is a fit for Python teams building long-running, approval-gated processes on Mistral models. One review lists it as a poor fit for teams that want a model-agnostic framework, non-Python developers, anyone who needs a stable GA API now, and buyers who need published per-agent or sandbox pricing. beri.net’s recommendation for regulated workloads is to pilot on data that could already go to a SaaS vendor and hold production until Mistral puts a price, a GA date and a self-hosted control plane in writing.

What to watch next

• Workflows status: whether Mistral’s documentation drops the public preview notice to match the GA description in launch coverage.
• GA for Managed Agents and the SDK: and how far the API changes before then.
• Pricing: itemised prices for hosted sandboxes and managed deployments, which are not yet published.
• Private cloud and on-premises: a confirmed timeline and terms for running the full platform inside customer infrastructure.
• Old Agents API: a removal date for the deprecated API.