Mistral Studio Agents vs Workflows: What to Build and Where to Deploy

Use a Mistral Studio agent when the next useful step depends on what it discovers. Use a workflow when the process has defined steps, approvals and recovery rules. Combine them when a bounded investigation sits inside a business process, and choose hosting separately from both.

Mistral's November 5 Studio expansion brings managed agents and apps alongside workflow, context and evaluation improvements. It makes more of the infrastructure available in one platform, but a team still has to decide which decisions belong to the model.

Start with the decision you cannot predefine

An incident investigation is a reasonable agent task. The next query depends on whether the first log points to a failed deployment, a slow dependency or a configuration problem. You can describe the goal and provide tools without pretending to know the entire path in advance.

By contrast, checking a form, routing it to an owner and waiting for approval usually has a known structure. Giving a model freedom to reinvent that sequence adds uncertainty without adding much value.

Your process	Starting point	What you must define
One model request with no orchestration	Direct SDK call	Input, output validation and errors
Investigation with changing next steps	Managed Agent	Allowed tools, boundaries and stopping conditions
Known steps with waits and branching	Workflow	State transitions, approvals and failure handling
Investigation followed by a controlled action	Agent inside a workflow	The boundary between recommendation and execution

The Managed Agents documentation describes a harness that runs the model-and-tool loop and records its work in sessions. Both Managed Agents and the Agents SDK are in public preview, so behavior, limits and APIs can change. An impressive demo is not a reason to ignore migration work in your rollout plan.

Put a bounded investigation inside the process

Consider a support escalation that needs an explanation before it can be assigned. An illustrative workflow could validate the ticket, ask an agent to investigate read-only logs, present the evidence to an operator, and route the approved summary to the right team.

The agent should return a finding with references and unresolved questions. It should not decide to grant a refund, alter a customer's plan or message someone just because those actions appear useful. Those decisions need explicit business rules and the appropriate permissions.

Mistral Workflows is also in public preview. It provides durable execution based on Temporal, with recorded steps, configurable retries and waits for human or external input. That makes it useful for a process that must survive longer than one application request.

Keep the irreversible part of a process easy to inspect. If a reviewer cannot tell which action follows approval, the system is too vague regardless of which orchestration product you chose. For a first version, a clear handoff to a person can be more useful than another autonomous tool.

Separate worker hosting from model hosting

Studio offers local development, self-hosted workers and managed deployments. The deployment guide makes an important distinction: running your workers on your own infrastructure does not mean the whole Studio platform runs on your premises. Model deployment is another decision again.

For example, your worker could run in your private cloud while Mistral orchestrates it and an API supplies model inference. Saying that this setup is simply “self-hosted AI” hides the data paths an infrastructure team actually needs to review.

Before choosing a mode, draw four boxes: the worker, the model endpoint, the session or workflow state, and the systems the tools call. Write down where each runs and which data crosses between them. Include traces and error logs, since they can contain information that never appears in the final answer.

Managed hosting is attractive when operating workers is the main burden. Self-hosted workers can make sense when they need private-network access or existing compute. Full on-premises Studio is a separate enterprise deployment discussion, not a switch implied by choosing a worker location.

Check managed-deployment eligibility before debugging code

The managed-workflow quickstart requires pay-as-you-go to be enabled on the organization or an Enterprise plan. A paid Pro or Team subscription alone does not qualify, and the Free plan cannot run managed deployments.

Its source integration currently supports repositories on github.com, including private repositories through the Mistral GitHub App. GitLab and GitHub Enterprise are not supported by that documented flow. Confirm repository access before spending an afternoon trying to fix a deployment configuration that cannot see the source.

The quickstart also warns that deployment names must be unique because matching names can receive each other's executions. Use names that distinguish the project and environment, and check the destination in the console before enabling a real trigger.

This is a deployment decision guide, not a tested setup walkthrough. Exact compute costs, quotas and enterprise terms should be checked for the selected organization and configuration. An inference price alone cannot establish the total cost of running an application.

Test the process at its uncomfortable points

For the support example, start with fabricated tickets and a test log source. Try an empty investigation result, a tool timeout, an operator rejecting the recommendation and a worker restart while approval is pending. Write down the expected state after each event before running the test.

Also check what an operator sees. A trace containing dozens of model calls may help an engineer debug, but the support reviewer needs the specific evidence behind the recommendation and a clear way to reject it. Those are different interfaces, even when they describe the same run.

My default would be the smallest workflow that gives the team a dependable process, with an agent added only where investigation genuinely needs flexibility. Studio can reduce the amount of runtime infrastructure you own. The remaining design work is deciding exactly how much judgment you are prepared to delegate.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help teams explain complex enterprise AI products through clear technical stories, useful demos and founder-led content.

Book a 30-minute call.