Mistral Studio managed agents: how to use them

Mistral Studio managed agents are AI agents you write in Python with Mistral's new Agents SDK, then run on your own machine, on your own servers, or on Mistral's cloud, with tracing and human approvals built in. To try one, you install the mistralai-agents package with uv, add a Mistral API key, and run a starter agent from the command line in a few minutes.

The thread that launched it

On November 5, @MistralAI opened a launch thread with "Le Chonk brings the brains. Mistral Studio gets it working." Le Chonk is the company's nickname for the trillion-parameter model we covered in our Mistral Large 4 post. The thread introduced managed agents and apps, plus additions to workflows, enterprise connections, observability and evaluations.

Two follow-ups in the thread show what Mistral thinks this is for. One on workflows describes processes that mix "fixed rules, agent reasoning and human approvals," and says Mistral's own product operations team saves about 35 hours a week routing feedback this way. The other, on apps, pitches "internal tools your IT team can approve," with hosting, login and access controls included. Its example is a staffing app that matches people to projects.

Mistral's announcement adds the agent features: skills that capture a team's procedures, sandboxes for running code, subagents that take parts of a task, and hooks. It names ABANCA, BNP Paribas, Capgemini and CMA CGM as Studio customers in production.

How to try it

The quickstart needs uv and an API key from console.mistral.ai.

1. Create a project: uvx --python 3.12 --from mistralai-agents agents init hello-agent, then cd hello-agent. The init step asks for your key and saves it to .env.
2. Run the sample: uv run --env-file .env python main.py. It's a weather agent with one tool, running on mistral-large-4.
3. Chat with it: uv run --env-file .env agents chat.
4. Add a second tool, such as the docs' get_forecast(city, days), and ask it something like "Will it rain in Paris in the next 3 days?"

The bare-bones version from the overview page is five lines of Python:

import asyncio
from mistralai import agents
agent = agents.Agent(model="mistral-large-4")
result = asyncio.run(agent.run("Say hello"))
print(result.text)

Three ways to run an agent

By default a session ends after 24 hours, or after 4 hours idle, and both limits can be raised. Tracing comes on with no setup: each turn becomes a trace in Studio, with one span per model call and per tool call. An agent has no code sandbox until you add one.

The default to change first

The setting worth checking is approvals. Mistral's approvals docs say the default permission_mode, "auto", follows your permission_policy, "or runs when the Agent has none." So an agent with no policy calls its tools without asking. Set "ask" on anything that sends, deletes or spends money. An approval waits one hour, and if no one answers, the call doesn't run, so a missed request fails safe.

What's live and what it costs

According to the release notes, Managed Agents, the workflow builder, the Apps CLI, tracing, webhooks and evaluations are all in public preview on Mistral's hosted cloud. HTTP API connectors are private preview. Only Studio apps in Vibe Code Web are generally available. The docs warn that the API "can change during the preview."

Mistral hasn't posted a separate price for managed agents. An October release note says managed deployments need a paid plan (Pay-as-you-go or Enterprise) during the preview and are subject to a quota. For a sense of model cost, the pricing page lists Mistral Large at $0.50 per million input tokens and $1.50 per million output. Some outside write-ups say private-cloud and on-premises versions arrive at the end of November. Mistral's docs don't give a date.

My read

This is Mistral catching up to the managed-agent products the bigger labs already ship, like Claude Managed Agents. The quickstart is short, and Mistral's pitch is where Studio can run: its cloud, your own VPC or on-premises. One catch for the privacy-minded: even in agents serve mode, on your own servers, Mistral stores the session. Until it leaves preview, I'd keep it to internal tools you can afford to rebuild.

What this means for businesses

The workflows post is the one owners should read. Fixed rules, an AI for the judgment calls and a person signing off on the risky step fit invoice routing or intake triage as well as they fit feedback sorting. Pick one process with a clear owner and write it down before you choose a platform. If you want help finding the right first one, New Face Design offers a free process audit.