Mistral Managed Agents Guide: Deploy, Modes, Quotas

Last updated: November 5, 2026

Mistral Managed Agents, listed in the November 5, 2026 Studio release notes, lets you build long-running agents once with the Agents SDK (mistralai-agents) and run them locally, self-hosted with agents serve, or Mistral-hosted with agents deploy. Deployed sessions are durable, support approvals, and send Studio traces by default.

Contents

What Managed Agents is
Three run modes
How to deploy with agents deploy
Session limits, hardening and tracing
Plan quotas for managed deployments
Limitations and comparisons
FAQ

This guide covers the Managed Agents public preview announced with the Studio expansion on November 5, 2026. Facts come from Mistral’s release notes, Managed Agents overview, run-modes and agents deploy docs, managed-deployment quotas, and the Studio news post. For the model often used in the hello-world examples, see our Mistral Large 4 open-weights guide.

What Managed Agents is

Mistral describes Managed Agents as autonomous agents that can work for hours or days, resume after a crash and run many sessions in parallel while you control tools, data and approvals. You define the agent in Python (model, instructions, tools, connectors, skills, environment and approvals). The Agents SDK runs the model and tool loop and records each step in a session. The overview’s hello world installs the SDK, sets MISTRAL_API_KEY and runs agents.Agent(model="mistral-large-4"). For one-shot completions Mistral still points to Chat completions; for fixed branching it points to Workflows. The same day’s Studio news post groups Managed Agents with managed apps, workflow builder features, HTTP connectors, evaluations and observability. If you mainly use the coding CLI, see our Mistral Vibe 3.1 CLI guide.

Three run modes

The run-modes documentation says the agent definition and session API stay the same in every mode. Only who runs the harness and where the session is stored change:

Mode	Start with	Who runs the harness	Session storage	Best for
Local, in-process	`agents.Agent(...) or agents chat`	Your program	In your process; ends with it	Development and tests
Self-hosted	`agents serve`	You, on your servers	Stored by Mistral; survives crashes	Production when tools must stay in your network
Mistral-managed	`agents deploy`	Mistral	Stored by Mistral; survives crashes	Production with nothing to host

Deployed modes keep each finished model and tool call in a durable session. If the worker dies mid-turn, another worker usually resumes from the last finished step. A tool that was still running is not replayed; about a minute later the model gets an error and may call it again, so side-effecting tools must be idempotent.

How to deploy with agents deploy

Use these steps when you want Mistral to host the worker:

Build locally with agents.Agent(...) or agents chat, and set an explicit name=.
Push the project to GitHub and connect the GitHub connector in Studio so the platform can clone it.
Use a MISTRAL_API_KEY with Workflows access (Developer, Developer Contributor, Workspace Contributor, Observability Viewer or Admin). A MistralSandbox also needs the cloud_user role.
Commit and run agents deploy. On a uv project without a Dockerfile, accept the generated Dockerfile and .python-version (3.12), commit them with pyproject.toml and uv.lock, then deploy again.
When the agent is live, try agents chat <name>, Vibe Code Web, or agents.Agent.load_from(agent_name=...).
Bind secrets with --secret-binding NAME=secret:workspace:... on every deploy (references only). A deploy replaces prior bindings.
On failure, use agents logs <name> --follow, fix the Dockerfile or missing modules, and redeploy.

Hardened deployments are the default when an agent acts for the starting user (for example through connectors). Do not pass --no-hardened; hardening lasts for the deployment’s life, so changes need agents shutdown and a new deploy.

Session limits, hardening and tracing

Default deployed limits are 4 hours idle and 24 hours total. Dockerfile ENV values can raise them to 24 hours idle and one week total; a week-long total also needs a higher AGENTS_SDK_EXECUTION_TIMEOUT_SECONDS. Human approvals wait one hour by default so a turn can pause while your app is closed.

Studio tracing is on by default for served and managed agents, with spans for agent, model and tool steps (including prompt and completion content unless disabled). Redaction runs on the worker first. Set AGENTS_SDK_STUDIO_TRACING_INCLUDE_CONTENT=0 or AGENTS_SDK_STUDIO_TRACING_ENABLED=0 in the Dockerfile for less detail. Organizations that disabled Studio observability get no traces.

Plan quotas for managed deployments

Mistral’s managed-deployment quotas page (documented for managed workers) sets concurrent caps per organization:

Plan	Concurrent managed deployments
Free	0
Pro or Team, pay-as-you-go off	0
Pay-as-you-go enabled	3
Enterprise	10

Managed deployments need pay-as-you-go or Enterprise. Creating or starting past the cap returns HTTP 403 until you stop or delete another deployment. Mistral Cloud picks instance size and does not expose per-deployment CPU or memory. Model tokens still bill at ordinary API rates; see our Mistral API pricing guide.

Limitations and comparisons

Managed Agents and the Agents SDK are in public preview: API, behavior and limits can change and may require code updates. Local sessions are not durable without the experimental journal, and interrupted tool calls are not safely replayed.

Compared with Microsoft’s Copilot harness in Foundry, Managed Agents keeps the agent in Python with an explicit self-hosted worker path. Compared with Open Duet 0.8, it is a first-party Studio product with managed hosting, tracing and workspace secrets. For speech agents, see the Voxtral 3 guide.

Demand signals on November 5, 2026: Google autocomplete returns “mistral managed agents” and “mistral agents sdk”; X recent-post counts for English posts matching mistral with “managed agents”, mistralai-agents, “agents deploy” or “Agents SDK” totaled 247 over about seven days (106 on November 4 UTC; 110 so far on November 5 UTC). No search-volume figures are available, and none are invented.

Frequently asked questions

What are Mistral Managed Agents?

Managed Agents is Mistral’s public-preview platform for autonomous agents built with the Agents SDK (mistralai-agents). You define the model, instructions, tools and environment once, then run the same code in your process, self-host with agents serve, or let Mistral host it with agents deploy. Deployed sessions are durable and can run for hours or days.

When were Mistral Managed Agents released?

Mistral listed Managed Agents in the Studio release notes dated November 5, 2026, as a new public-preview, Mistral-hosted SDK release. The same day’s Studio news post describes the expansion of Studio with managed agents and apps.

What is the difference between agents serve and agents deploy?

Both store sessions on Mistral’s control plane so work can outlive a crash. With agents serve you run the worker on your own servers so tool code stays in your network. With agents deploy, Mistral builds your GitHub project and runs the worker for you, so you host nothing.

Do I need pay-as-you-go for managed deployments?

Yes for Mistral-managed workers. Mistral’s quotas page says Free plans and Pro or Team plans without pay-as-you-go have a concurrent managed-deployment cap of 0. Pay-as-you-go allows 3 concurrent deployments; Enterprise allows 10. Creating a deployment beyond the cap fails with HTTP 403.

How long can a Managed Agents session run?

By default a deployed session ends after 4 hours idle or 24 hours total. The agents deploy docs say you can raise those with Dockerfile ENV settings up to 24 hours idle and one week total (and you may also need to raise AGENTS_SDK_EXECUTION_TIMEOUT_SECONDS).

Is Studio tracing on by default for Managed Agents?

Yes for self-hosted and Mistral-managed modes. The overview and deploy docs say each turn becomes a Studio trace with one span per model call and tool call, with no extra package. Spans are redacted on the worker before they leave. Local in-process runs do not send Studio traces.

Sources:

Mistral Managed Agents: Power-User Guide to Deploy and Run Modes