Mistral Managed Agents is a new way to build an AI agent in Python once and then run it on your laptop, your own servers, or Mistral’s cloud. Mistral announced it on November 5, 2026 as a public preview. The big draw: deployed agents keep working in the background for hours or days, and they pick up where they left off after a crash.
Here are the facts worth knowing before you read on.
- Announced November 5, 2026, as a public preview on Mistral-hosted infrastructure.
- Built with the Python Agents SDK, installed as
mistralai-agents. - Three ways to run: in your own process, with
agents serve, or withagents deploy. - The release notes don’t mention pricing or plan requirements.
If you’ve been wiring agents together with your own queue and database, this one is aimed at you.
How Mistral Managed Agents work: one codebase, three places to run it
You define the model, instructions, tools and environment in one project. Then you choose where it runs. The code stays the same, which is the whole pitch of “write once, run anywhere.”
Running it in your own process is the easy path while you build. Self-hosting and Mistral’s cloud are for when the agent needs to run without you watching.
| Run mode | Command | Where it runs | Best for |
|---|---|---|---|
| In your process | Run your script | Your machine | Building and testing |
| Self-hosted worker | agents serve |
Your own servers | Tools that must stay inside your network |
| Mistral-managed | agents deploy |
Mistral’s infrastructure | Long jobs you don’t want to host |
Even when you self-host, Mistral still stores the sessions. You don’t have to run a database, but your session data does live with Mistral.
What happens when an agent crashes or needs your approval
Deployed agents run in the background and can work for hours or days. If one crashes, it resumes from the last completed step instead of starting over. That matters for long jobs, where a restart can waste real money.
You also get a few controls over a running agent:
- Require a person’s approval before specific tool calls run.
- Steer or interrupt a turn while it’s in progress.
- Reopen a session after your own app restarts.
The approval step is the one to use first. A turn can pause for a human and then resume later, so risky actions like sending email or deleting files don’t need to happen unattended.
Each turn also shows up as a trace in Studio, with one span per model call and tool call. There’s no setup for that.
How to deploy your first agent with Mistral Managed Agents
Mistral’s docs use a Quickstart project called hello-agent. Before you deploy, you need three things in place.
- A GitHub repository for the project, with the GitHub connector linked to your account in Studio.
- An API key whose user has a workspace role with Workflows access, such as Developer or Workspace Contributor.
- The project scaffolded with
agents init.
An admin assigns that workspace role in the admin console, so ask early if you’re not the admin.
Then push your code and deploy it:
git add --all && git commit -m "Add get_forecast"
git push -u origin HEAD
uv run --env-file .env agents deployDeploys use the pushed commit, not the files on your laptop. That trips people up, so push first. Per Mistral’s deploy guide, the command builds the Dockerfile from agents init and waits until the agent is live.
After that, agents chat hello-agent lets you talk to it and agents logs hello-agent --follow shows what it’s doing. Your own app can load it by name with Agent.load_from(agent_name="hello-agent").
What Mistral hasn’t said yet
This is a public preview, so expect changes. The release notes don’t cover pricing, quotas or usage limits. Check your account settings before you point an agent at a big job.
One more detail comes from a secondary source. According to AI Weekly’s November 6 edition, private cloud and on-prem deployment arrive in late November. Treat that date as reported, not confirmed by Mistral’s docs.
The idea of hosted agents with memory and schedules isn’t unique to Mistral. AWS is pushing a similar approach, and we covered its credit bill to watch in our Kiro Bots post. If you want a free, open-source option instead, see our look at the DeepSeek Super-agent.
Frequently Asked Questions
What is Mistral Managed Agents?
It’s a Python SDK and hosting service from Mistral for building autonomous agents. You write the agent once, then run it locally, on your own worker, or on Mistral’s servers.
How much does Mistral Managed Agents cost?
Mistral hasn’t published pricing in the announcement or the deploy guide. Check Mistral’s pricing page or your Studio account before heavy use.
Do I need GitHub to deploy an agent?
Yes, for agents deploy. Mistral’s guide requires a GitHub repository and the GitHub connector in Studio. Self-hosting with agents serve is the alternative.
What if my agent shows as “not observed”?
It means nothing is running it. Deploy again, or start agents serve on your own machine.
Should you try it now?
If you already build agents in Python, yes, on a small side project. The crash recovery and approval pauses solve problems you’d otherwise build yourself. But hold off on anything important until Mistral shares pricing and the preview settles down.
Leave a Reply