AI agent or workflow: what SMEs should check
On November 5, 2026, Mistral added a public preview to Studio that matches a question operations groups ask first: does the task need an AI agent or a workflow? Managed Agents and the workflow builder shipped the same day. The agent investigates and acts. The workflow decides which steps follow fixed rules, where a model judges, and where a person decides.
For an SME this is not a switch to flip on Monday. The preview can change, and Lyron has not tested it in a customer operation. It still frames the decision that comes before the first production agent: what should stay fixed in the process, and what may the model choose case by case? How an AI agent works in general is on our page about AI agents. The wider frame is on the page about process automation.
This article reads the announcement and the Mistral documentation. Checked on November 15, 2026. Every technical detail comes from Mistral.
In short
- Public preview since November 5, 2026: Managed Agents in the Agents SDK, and a workflow builder that turns a description into a process, tests it, and deploys it. You can read and edit the generated code.
- Agent: The model chooses the next step and can work for hours or days. Once deployed, it resumes after a crash from the last finished model or tool call. Mistral stores the session.
- Workflow: The steps live in code. A run can wait for a person or an event and resume from the last completed step. The documentation describes runtimes from seconds to months.
- Separate places: For a workflow, Mistral hosts the orchestration and your worker runs the code. For a deployed agent, Mistral stores the session even when the harness runs on your servers.
- For most SMEs: A narrow process with fixed steps and an approval before anything is written or sent remains the closer first step. Add an agent when the path changes from case to case.
What shipped on November 5
The release notes list a set of Studio pieces for that day, almost all in public preview and hosted by Mistral. Four of them decide the agent-or-workflow question directly.
- Managed Agents. You describe the model, instructions, tools, and environment once in Python. The same code runs in your process, on a worker you start, or in a deployment Mistral operates. A running task survives a crash and can wait for an approval.
- Workflow builder. You describe the process in plain language. Vibe writes the workflow, tests it, and can deploy it. You can inspect and edit the generated code at every step. Mistral describes that code as something you can read, not a closed box.
- Webhooks and triggers. An event in a connected tool starts a workflow or an agent session, instead of waiting on a schedule. The note names built-in ingestion for services such as Notion and Lucca, including identity resolution, normalization, and deduplication.
- Traces and evaluations. Each turn of a managed agent writes a trace of the agent, the model call, and the tool call, with no extra package. Common secrets and personal data are redacted on the worker before the trace leaves. Offline evaluations run against your own datasets. Online evaluations score live sessions. You can turn a trace into an evaluation record.
The same day, a private preview added a way to register a web API as a connector: you describe its operations once, and agents and workflows call them after that. You do not have to run a separate adapter for that API. Once a process is running, the article on monitoring AI workflows covers how stops and samples show up in day-to-day operations.
Three roles in one process
The announcement separates three roles that can sit in the same process. An agent investigates and acts. A workflow coordinates execution and approvals. An app gives employees a way to use both. Studio puts them on a shared base: a model, a connection to your data and systems, the runtime, and a way to look up what happened afterward.
Mistral writes about its own product operations team. A workflow there sorts ideas from inside the company and feedback from customers. A model identifies bugs and feature requests, and rules route them to the right team. Mistral states a saving of about 35 hours a week. That is Mistral’s figure for its own team, not a result you can carry over to your shop.
The announcement names companies that already use Studio in production, including ABANCA, BNP Paribas, Capgemini, and CMA CGM. It does not spell out, case by case, what those companies automate with it. For an SME, the split between the roles matters more than the list of names.
What the agent does, and where the session lives
The Managed Agents documentation describes one turn like this. You send a message to a session. The harness passes the conversation to the model. The model answers or picks a tool. The harness runs the tool and sends the result back, until an answer stands. A session is the record: every message, every tool call, every result.
The model runs on Mistral’s API or behind a gateway you choose. The harness, the loop of model and tool, can sit in three places:
| What you want | How you start | Who runs the harness | Session |
|---|---|---|---|
| Develop and test | In your own program | You | Lives in the process and ends with it |
| Production, tools inside your network | agents serve |
You, on your servers | Mistral stores it, and it survives a crash |
| Production, nothing to host | agents deploy |
Mistral | Mistral stores it, and it survives a crash |
Once you deploy, a few limits matter before any test. The documentation says a session ends after 24 hours by default, or after 4 hours idle, when no turn is running or the turn is only waiting for you. You can raise both limits. An approval waits one hour by default, even while your app is closed. You choose which tool calls need approval, and whether your code or a person answers. You can steer a running turn with a message, or interrupt it.
Managed Agents and the SDK are in public preview. The API, behavior, and limits can change during the preview, and an update can require changes to your code. The documentation shows the approval pattern on a payment-review agent: the agent reads the material on its own and waits for a person before it opens a case. The connected systems in that example are faked.
One boundary is easy to mix up with a different announcement. When a company runs Claude’s agent service itself, the session sits on infrastructure that company operates. The article on self-hosting AI agents covers that split. With Mistral Managed Agents, Mistral stores the session once you deploy, including when the harness runs on your servers with agents serve. A harness inside your network does not mean the record stays there.
What a workflow keeps fixed
The workflow documentation describes workflows as multi-step processes that combine model calls, tools, external APIs, and human input. You write the process as code. The platform handles execution, retries, schedules, and the connection to the rest of Mistral. Workflows are in public preview as well. Mistral says it does not plan major changes to APIs and features, that they might still happen, and that it will give notice before they do.
The agent page puts the choice in one line. A single model call without a loop stays an ordinary API call. A process with explicit steps, retries, and branching is a workflow.
Four points from the documentation decide whether a workflow is the better frame:
- Every step is in a history. If the process dies, another one resumes from the last completed step. Mistral names Temporal as the engine for that durable execution.
- Retries belong to the step. Anything that touches the outside world, a model call, an HTTP request, or a write, lives in an activity and can run again with backoff. The workflow only coordinates, and it has to stay deterministic so the platform can replay it from history.
- Waiting is built in. A run pauses on a human signal or an external event and continues when the input arrives. The documentation describes runtimes from seconds to months.
- The place is hybrid. Mistral hosts the orchestrator: state, history, and task dispatch. Your code and the workers run with you, on your own machine while you develop, or in your own infrastructure in production. Workers connect outbound. The orchestrator does not open a connection into your network. For enterprise customers, the documentation says the Studio environment can also run in a private cloud or on premises.
The same page names two paths that make the data flow concrete. The SDK can encrypt payloads before they leave your worker. The platform then stores ciphertext only. Payloads above 2 MB can be offloaded to storage you operate. The platform keeps the reference. Inputs and outputs still pass through the platform. Whether this split fits your contracts and your processing terms is something you settle for your own operation. This article is practical guidance and does not replace legal advice.
| Question | Managed agent | Workflow |
|---|---|---|
| Who chooses the next step | The model, inside the loop | The process you wrote |
| Waiting for a person | Approval, one hour by default | A signal or an event, from seconds to months |
| After a crash | Last finished model or tool call, once deployed | Last completed step |
| Where the state lives | Mistral stores the session once you deploy | Mistral holds the orchestration, your worker runs the code |
A description in, code you can read
The builder from November 5 lowers the starting effort. It does not lower the duty to read what comes out. You describe the process, Vibe generates the workflow, tests it, and can hand it to a managed deployment. While that happens, the graph of the process can sit beside the conversation. You can read the code at every step, and you can change it. That is the difference from a process that exists only on a screen you cannot open.
Managed deployment is older than the builder. The release notes of October 9, 2026 say that Mistral runs the workers, that you do not stand up your own server stack for them, and that these deployments in preview require a paid plan and sit under a quota. The documentation describes the path through a GitHub repository: Mistral builds an image from it and runs the container. You do not create an API key for the worker. The preview and its limits can change. The page does not name a price in euros, so this article does not invent one.
For an SME the order is plain. Write the process so another person can follow it. Then check that the generated code contains those steps. Only then connect a real case. A generator does not replace the approval. How tightly to instruct a model when it only drafts one sentence inside such a process is in the guide to instructing AI agents reliably.
Example scenario from an SME
Example scenario, not a customer case: A machinery supplier in North Rhine-Westphalia receives emails about delivery dates. The path is almost always the same. The order number is in the message, the status is in the order system, someone writes a short sentence, and a person approves it before it goes out.
In the logic of the documentation, that is a workflow, not an agent. The steps are known: accept the message, read the number, fetch the status, write a draft, wait for approval, then send. The model writes only the sentence, and only when the status is more than a date. The approval sits before the send. An agent is the later look, when the request changes from case to case, for example because one email mixes three open orders and the path cannot be fixed in advance. Even then, sending waits for a person. The scenario does not carry over the 35 hours Mistral states for its own team. How an approval like that is bounded in practice is in the article on permissions and approvals.
When a fixed workflow is the closer step
Most processes in a mid-sized company do not need an agent that chooses the next step unsupervised. They need a process someone can read: fetch the data, let the model do one bounded step, check the result, write it on. A workflow in n8n can be kept that narrow and handed to the team. Mistral’s preview is another place where the same split now exists as a product. A new preview is a weak reason to make it the first step.
An AI agent earns its place when the path changes from case to case and fixed steps are not enough. Limits and a human approval still come first. The preview is the right lens when three things come together: the path is not the same every time, someone reads the generated code and the session record, and it is acceptable for your operation that the orchestration or the stored session sits with Mistral, unless you use the paths the documentation describes, such as your own gateway, the worker inside your network, or encryption of the payloads. If one of those is missing, a narrow workflow is the more practical first step.
What to settle before a test
- Write the process down: which steps are fixed, where a model may draft, and where a person decides.
- Start with the workflow when the path is known. Add the agent when fixed steps are not enough.
- Note the data flow: where the model call goes, who stores the session, who holds the orchestration, and where your worker runs.
- Approval before anything takes effect: writing, sending, and paying wait for a person. The documentation names one hour as the default wait on the agent.
- Know the session limits: 24 hours and 4 hours idle are the defaults, and both can be raised. Limits can change during the preview.
- Read the generated code before it touches a real order. The builder returns text you can check.
- Start with one narrow case: one process, one draft, one person approves. Widen it when the process holds up in daily work.
Sources
Checked on November 15, 2026, published on November 17, 2026. Technical details come from Mistral. The reading and the example scenario are Lyron’s.
- Mistral release notes – entries of November 5, 2026 on Managed Agents, the workflow builder, triggers, traces, and evaluations, and the October 9, 2026 entry on managed deployments.
- Mistral: Build Enterprise Agents, Workflows & Apps in Mistral Studio – the roles of agent, workflow, and app, and the figure of about 35 hours a week for Mistral’s own product operations team.
- Managed Agents – the three run modes, session limits, the approval wait, and where the model runs.
- What is Mistral Workflows? – history, waiting for a person, hybrid orchestration, encryption, and offloading payloads above 2 MB.
- Managed deployments – Mistral runs the worker from a repository. The preview and its limits can change.
Want a process fixed before a model chooses the next step?
Lyron sets up a narrow first process with you: fixed steps, the model writes only where the path is not clear, and a person approves before anything goes out.
Book a free consultation