mistral studio agents: Managed Agents Hit Public Preview

The mistral studio agents release is Mistral's bid to move enterprise agent building from bespoke infrastructure to a managed platform. On November 5, Mistral announced that Mistral Studio, its enterprise AI platform, now includes managed agents and apps, alongside new capabilities for workflows, enterprise context, agent observability, and evaluations. The Agents SDK and Managed Agents runtime are in public preview, and Mistral is explicit that APIs, behavior, and limits can change during the preview. Source

What Mistral shipped

The release has several moving parts. Managed Agents is the runtime for autonomous agents that work on their own for hours or days, pick up where they stopped after a crash, and run many sessions in parallel. The Agents SDK, distributed as the Python package mistralai-agents, drives the model and tool loop and records every step in a session. Workflows got event triggers from enterprise systems and a visual graph for inspecting logic. And Studio apps now come with hosting, authentication, and access controls, so a coding agent's output can be put in front of employees with enterprise requirements attached. Source

Underneath it all sits Mistral Large 4, which Mistral has nicknamed Le Chonk, positioned for coding, defensive cybersecurity, and reasoning over complex documents. Studio connects the model to enterprise data and systems through connectors and MCP servers, and a new Mistral MCP server lets developers build with Mistral Vibe or their existing coding agent, test locally, and deploy to Studio with one command. Source

How Managed Agents works

The conceptual model in the docs is clean. An agent is a model plus a harness. The model decides the next step: answer, or call a tool. The harness runs the loop: it calls the model, runs the tools the model picks, and records each step. A session is the record of what happened, one entry per message, tool call, and result, grouped into turns. An environment is the computer the agent works on, a sandbox where it runs shell commands and edits files; by default an agent has none. Source

The run modes form a ladder. On your laptop, the harness and your tools run in your Python process, the session lives in memory, and any sandbox is a local Docker container. One step up, the sandbox moves to Mistral while everything else stays local. In production you deploy with Studio two ways: agents serve runs the harness and your tools on your own servers with Mistral storing the session, and agents deploy runs everything on Mistral so you host nothing. Source

Durable execution is the headline operational feature. A task can run for days; if the machine running it crashes, the turn continues from its last finished model call or tool call, and your application can exit and come back for the result. By default a session ends after 24 hours, or after 4 hours idle when no turn is running, and both limits can be raised. Approvals work the same way: when the agent hits a tool call that needs a person's sign-off, the turn waits for the answer, one hour by default, even while your app is closed. Source

The turn cycle is worth understanding because it defines what a session can record. You send a message to a session; the harness sends the conversation to the model; the model answers or picks a tool; the harness runs the tool and returns the result to the model, repeating until the model answers; the answer comes back as the turn's result. A message sent while a turn runs joins that turn rather than queueing behind it. The docs walk through a bank payment review example: an alert about a suspicious transfer opens one session, the agent reads the account history, KYC profile, sanctions lists, and news on its own, then waits for an analyst's approval before opening a case, with every step auditable in the session afterward. A duplicate alert finds the same session instead of starting a second review. Source

Observability comes built in. Each turn is a trace in Mistral Studio with one span per model call and tool call, needing no setup, and the announcement describes OpenTelemetry traces showing each step of a run, including model calls, tool use, and approvals. Evaluations let teams test changes to prompts, models, or tools against their own datasets before release. Source

The agent shape: skills, sandboxes, subagents, hooks

The announcement gives developers four composition tools. Skills capture a team's procedures for the agent to follow. Sandboxes let the agent run code. Subagents let it delegate parts of a task. Hooks add custom logic at key points in execution, which is the integration seam for existing systems. Developers set boundaries on the tools, data, and actions the agent can access, and Studio manages the runtime, which Mistral says is the same infrastructure behind millions of Mistral Vibe sessions each week. Source

Mistral's own engineering dogfood is an incident investigation agent: when an alert fires, the agent follows the team's runbook, asks for approval before sensitive actions, and posts findings in Slack, which Mistral says saves the team hundreds of hours each month. Its product operations team built a workflow that classifies incoming ideas and feedback and routes them, saving about 35 hours a week. Its Applied AI team built a staffing app where an agent recommends people for projects and explains the fit. These are Mistral's own reported figures. Source

The sovereignty pitch

Studio's enterprise framing is control over where AI runs. Models and platform can deploy in your own cloud, VPC, or on-premises environment, in Mistral-hosted deployments, or a combination. Access to data and tools is governed by the customer, and the customer decides which actions require human approval. Traces and evaluations remain the customer's, which Mistral frames as preserving operational knowledge. Source

The named production users are heavyweight: ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM, and Tata Consultancy Services. BNP Paribas's quote in the announcement is direct about why: the AI systems they build run sensitive processes, so they run Studio on their own infrastructure. CMA CGM says processes built with Mistral's engineers reach production readiness in weeks rather than months. These are vendor-published customer statements, so read them as positioning, but the deployment model they describe is a real differentiator for regulated European enterprises. Source

Our read

This section is our assessment, not Mistral's claims. The interesting competition here is not model against model; it is runtime against runtime. Anthropic's Managed Agents and now Mistral's Managed Agents are converging on the same shape: a hosted loop with durable sessions, approval gates, sandboxes, and traces, so the team's code shrinks to tools, instructions, and boundaries. We covered the memory-maintenance side of this pattern in claude dreams, and the long-running autonomous teammate pattern in perplexity teammates.

Two cautions before teams jump in. First, public preview means public preview: Mistral says plainly that the API, behavior, and limits can change and that an update can require code changes, so treat this as evaluation-stage infrastructure, not a foundation for a compliance-critical process this quarter. Second, an agent that reads your alerts, account histories, and documents is only as safe as its approval gates and tool boundaries; our piece on AI agent memory poisoning covers what happens when untrusted content reaches agent state, and the mitigation is the same discipline Mistral's docs push: approvals on sensitive actions, least-privilege tools, and auditable sessions.

For European enterprises with data residency requirements, Mistral Studio is now the most complete self-hosted agent platform story from a frontier lab. For everyone else, the practical takeaway is simpler: the managed-runtime pattern is becoming table stakes, and the differentiator is shifting to your tools, your runbooks as skills, and your evaluation datasets.