Mistral Studio opens managed agents in public preview

• Managed Agents and Apps are in public preview on Mistral Cloud; Workflows is generally available, according to the rollout announcement.
• Feedback routing saves Mistral's product operations team about 35 hours a week, the company says in its workflow demo.
• Hosted apps include authentication and access controls, as the app announcement describes.
• Private-cloud and on-premises availability for these releases is scheduled for the end of November, per the availability post.

A self-hosted agent worker can still store its sessions with Mistral. An app deployment can partially succeed and return an error, while Studio can turn production trace spans into evaluation records.

The mistralai-agents Python SDK runs the same agent definition in three modes:

• In-process: the harness, tools and session live in your application.
• Self-hosted: agents serve runs the harness and tools on your servers, with sessions stored separately by Mistral.
• Mistral-managed: agents deploy moves the harness and tools onto Mistral's infrastructure.

The best addition is crash recovery: deployed turns resume from their last completed model or tool call. They can continue after the calling application exits.

Mistral's launch post describes the configurable harness components:

• Skills: reusable team procedures.
• Subagents: delegation of parts of a task.
• Sandboxes: environments for running code.
• Hooks: custom logic at execution points.
• Approvals: gates before sensitive actions.

Mistral's engineering team built an incident-investigation agent that follows runbooks, requests approval for sensitive actions and posts findings in Slack. The company says it saves hundreds of engineering hours each month.

Sessions default to a 24-hour lifetime or four hours idle; both limits can be raised. Approval waits default to one hour, and the SDK documentation warns that preview updates can require code changes.

Workflows combine deterministic steps, agent reasoning and human decisions. Interrupted executions can resume where they stopped.

The November 5 release notes distinguish the GA workflow system from the natural-language Workflow builder, which remains in public preview.

• Authoring: Vibe generates, tests and deploys workflows from a description, leaving editable code.
• Inspection: a read-only graph renders activities and execution paths from the workflow code.
• Triggers: external events can start workflows or agent sessions through Studio or the API.
• Event handling: built-in ingestion for services including Notion and Lucca handles identity resolution, message normalization and deduplication.

Mistral's feedback-routing example uses a model to identify bugs and feature requests, then applies rules to send submissions to the appropriate team.

Studio provides hosting, authentication and access controls, plus reusable, pre-approved components intended to reduce security-review work. Mistral's staffing app uses an agent to recommend people for projects and explain the supporting experience.

The documented deployment path has several concrete constraints:

• Build input: standard mistral deploy builds a pushed GitHub commit, using modules and Dockerfiles declared in mistral.json; --from-disk provides an alternative.
• Environment: the command deploys app modules into production, with no flag for another environment.
• Worker: an app can have one workflow worker. Its deployment runs alongside other modules, so a worker failure can leave those modules deployed while the command exits with an error.

The release notes add two ways to supply enterprise context:

• HTTP API connectors: register a web API's operations without building and hosting an MCP server. Agents can select operations, while workflows can call them programmatically.
• Managed search indexes: provision a hosted vector index from a name, embedding model and optional schema. Ingestion and search use the Mistral SDK, with compatibility across the Search Toolkit's pipelines, QueryEngine and retrievers.

The search service manages the backing storage rather than requiring a separate search backend deployment.

Managed agents emit agent, model and tool spans to Studio by default. According to the observability release notes, common secrets and personal data are masked on the worker before spans leave it; custom redaction policies are also available.

The evaluation additions cover three distinct surfaces:

• Offline runs: mistralai-evaluations and the workflow evaluation plugin execute evaluations from application code and store results in Studio for version comparison.
• Managed datasets: CSV, JSONL and free-format records support test data beyond conversations.
• Online runs: evaluations score incoming production traces, with a UI for testing an evaluator against live traces before enabling it.

Mistral says customers can deploy both models and the Studio platform in their own environments. Its Applied AI team is building customer systems ranging from penetration testing to KYC onboarding.

The new releases reach private-cloud and on-premises deployments at the end of November; launch-day availability is on Mistral Cloud.

Public preview also has billing gates. The app deployment documentation lists these organization-wide quotas:

• Free: zero apps.
• Pro or Team without pay-as-you-go: zero apps.
• Pay-as-you-go enabled: three apps.
• Enterprise: ten apps.

Apps across all workspaces count toward the organization's quota. Each app can have three frontend/backend modules combined and one managed Postgres 17 database.