Mistral Studio adds managed agents and apps in public preview
Mistral AI has expanded Mistral Studio, its platform for building and running enterprise AI systems, with managed agents, a natural-language workflow builder, a command-line tool for deploying apps and a new set of evaluation and tracing features. The update is dated November 5, 2026 in the Mistral documentation release notes, which list twelve new entries: ten in public preview, one in private preview and one generally available. Every entry in that list is marked as Mistral-hosted.
The centerpiece is a new Agents SDK, distributed as the Python package `mistralai-agents`, which lets developers define an agent in code and then run it in three ways: inside their own process, on their own servers with `agents serve`, or on Mistral's infrastructure with `agents deploy`. Deployed agents keep running after the calling application exits, resume from their last completed step after a crash and can pause for human approval. Mistral describes the platform as built around Mistral Large 4, the open-weight model it nicknames "Le Chonk", and the documentation's first code example calls `mistral-large-4` directly.
The release also changes the recommended path for existing Mistral developers. The Studio documentation now files the older Agents API guides under a Deprecated section, and the page for durable agents inside Workflows carries the warning that they "are about to be deprecated", with new agents to be built on the Agents SDK. Mistral has published no removal date for either.
Pricing for the new features is not stated on the announcement, the release notes or the product page. The documented access rules are narrower: managed deployments of workflows require pay-as-you-go billing or an Enterprise plan, and the Studio overview says traces and evaluations are available to Enterprise-tier organizations.
··········
MISTRAL STUDIO NOVEMBER 5 RELEASE: WHAT SHIPPED AND IN WHICH STATUS.
The release notes give each new capability a status label and a hosting label. Public preview in Mistral's terms means the feature can be used now, while its API, behavior and limits may change, and an update may require code changes on the customer side. The table below reproduces the twelve entries as Mistral lists them.
........
Capability | Area | Status | Hosting |
Managed Agents | SDK | Public preview | Mistral-hosted |
Workflow builder | Studio | Public preview | Mistral-hosted |
HTTP API connectors | Studio | Private preview | Mistral-hosted |
Mistral Apps CLI | Studio | Public preview | Mistral-hosted |
Offline evaluations (Evaluation SDK) | Studio | Public preview | Mistral-hosted |
Managed datasets for evaluations | Studio | Public preview | Mistral-hosted |
Datasets from production telemetry | Studio | Public preview | Mistral-hosted |
Managed search index | Studio | Public preview | Mistral-hosted |
Tracing for managed agents | Studio | Public preview | Mistral-hosted |
Webhooks and triggers | Studio | Public preview | Mistral-hosted |
Online evaluations | Studio | Public preview | Mistral-hosted |
Studio apps in Vibe Code Web | Studio | Generally available | Mistral-hosted |
........
The only generally available item is the rendering of Studio apps inside Vibe Code Web, Mistral's browser-based coding environment, where apps appear through an MCP apps server and a live workflow graph sits next to the builder conversation. Everything else in the release, including the agents runtime that gives the update its headline, carries the preview caveat. Companies planning production use should read the labels as a signal that interfaces can still move before general availability.
··········
HOW MISTRAL STUDIO MANAGED AGENTS WORK.
Mistral's documentation defines an agent as a model plus a harness. The model decides the next step, either answering or calling a tool, while the harness runs the loop, executes tool calls and records each step. Conversations are stored as sessions made of turns, where a turn runs from the user's message to the agent's answer. An agent has no execution environment by default; when one is attached, it is a sandbox in which the agent can run shell commands and edit files, and it can run on Mistral, on the machine hosting the harness or on hosts the customer manages. Tools, MCP servers and connectors are declared on the agent and called by the harness when the model selects them.
The minimal example in the official documentation shows how little code a first agent needs:
import asyncio
from mistralai import agents
agent = agents.Agent(model="mistral-large-4")
result = asyncio.run(agent.run("Say hello"))
print(result.text)The same agent definition can then be run in three modes, which differ in where the harness runs and how long the session survives.
........
Run mode | Command | Harness runs on | Session storage |
In-process | agents.Agent(...) or agents chat | The developer's program | In memory, ends with the process |
Self-hosted | agents serve | The customer's servers | Stored by Mistral, survives crashes |
Mistral-managed | agents deploy | Mistral | Stored by Mistral, survives crashes |
........
In the two persistent modes Mistral stores the session, so the customer does not have to host a database for agent state. If the machine running an agent fails, the turn resumes from its last finished model call or tool call. A running turn can receive an additional message or be interrupted, and developers choose which tool calls require approval, which can come from code or from a person. Approvals outlive the calling application: a turn waiting for a decision holds by default for one hour even if the application that started it has closed.
The documentation sets default lifetimes for these long-running objects, all of which Mistral says can be raised.
........
Managed agent limit | Default value |
Maximum session lifetime | 24 hours |
Idle timeout (no turn running, or waiting for the user) | 4 hours |
Wait for a human approval | 1 hour |
........
The announcement on the Mistral website lists further building blocks for agents: skills that capture a team's procedures, sandboxes for running code, subagents for delegating work and hooks for inserting custom logic, with developers able to set limits on which tools, data and actions an agent can reach. The release notes add that deployed agents can run for hours or days. The 24-hour default session lifetime means that multi-day runs depend on raising that limit, a setting the documentation mentions without giving a maximum.
··········
MISTRAL STUDIO WORKFLOWS, CONNECTORS AND APPS.
Workflows are the deterministic counterpart to agents. Mistral's announcement describes them as processes in which the developer decides which steps follow fixed rules, where an agent reasons and where a person makes the decision, with interrupted runs able to resume. The new workflow builder lets a developer describe an automation in plain language inside Mistral Vibe, which generates, tests and deploys the workflow while exposing the generated code for inspection and editing at each step. New webhooks and triggers, defined in Studio or through the API, can start workflows, spaces or agents from external events, and the release notes name Notion and Lucca among the services with built-in ingestion.
For connecting enterprise systems, the private-preview HTTP API connectors let a team register a web API by describing its operations, without first building an MCP server around it. Mistral says the full set of operations stays visible, with no wrapper hiding fields or endpoints. A managed search index, also new, provisions a hosted vector index from a name, an embedding model and an optional schema, and works with the ingestion pipelines, query engine and retrievers of Mistral's Search Toolkit.
Apps are the user-facing layer. The Mistral Apps CLI scaffolds an app with `mistral apps init`, runs it locally with `mistral apps dev`, deploys it under Apps in Studio with `mistral apps deploy` and adds capabilities to an existing app with `mistral apps capability add`. According to the announcement, Studio supplies hosting, authentication and access controls for these apps, along with pre-approved reusable components, and apps can call agents, workflows, search indexes and shared enterprise context. The Studio documentation states that apps run on Mistral Cloud, while agents and workflows can run either on Mistral's cloud or on the customer's own infrastructure.
··········
TRACING AND EVALUATIONS IN MISTRAL STUDIO.
Tracing for managed agents is on by default. Agent, model and tool spans are sent to Studio without setup, and each turn appears as a trace with one span per model call and tool call. Before spans leave the worker, a default redaction policy masks common secrets and personal data, and customers can define their own redaction policies. The announcement adds that traces follow the OpenTelemetry format and show approvals alongside model calls and tool use.
The evaluation features are designed to catch regressions before and after a change ships. Offline evaluations run from the `mistralai-evaluations` SDK or a workflow evaluation plugin, with results stored in Studio for comparison across versions. Datasets are no longer tied to conversations and can be created from CSV or JSONL uploads, from free-format records, or from production traces, either one span at a time in the Trace Explorer or in bulk, with records open to review and annotation before use. Online evaluations run against live production traffic and can be tested on live traces in the interface before they are switched on.
··········
MISTRAL STUDIO ACCESS, PLANS AND DEPLOYMENT OPTIONS.
Mistral has not published prices for managed agents, apps, the workflow builder or the new evaluation features, and the documentation pages fetched on November 6 give no per-run or per-hour figures. Customers therefore have to read the cost of a deployment from model usage and their contract until Mistral publishes more. The access rules that are documented are summarized below.
........
Item | What Mistral documents |
Studio access | The pricing page FAQ says the Free plan includes access to Mistral Studio |
Managed deployments of workflows | Require pay-as-you-go billing or an Enterprise plan; the Free plan cannot run them |
Pro and Team plans | Do not qualify on their own; pay-as-you-go must be enabled on the organization |
Traces and evaluations | Listed in the Studio overview as available to Enterprise-tier organizations |
Concurrent deployments | Capped per plan, with figures on the Quotas page |
Python SDK | Python 3.12 to 3.14, installed with uv |
Deployment builds | Only from repositories on github.com; GitLab and GitHub Enterprise not supported |
Feature prices | Not published for the November 5 features |
........
Deployment location is the area where Mistral's statements need careful reading. The announcement says customers can deploy "in the cloud, in your own VPC, on-premises or across a combination of environments", and that both the models and the platform can run inside the customer's own environment. The November 5 release notes, however, label all twelve new entries as Mistral-hosted and give no date for self-hosted versions of these specific features. Customer quotes in the announcement describe existing self-managed setups: BNP Paribas says it runs Mistral Studio on its own infrastructure, and ABANCA says it runs Studio in its own cloud environment with inference provided by Mistral. Within the new Agents SDK, the `agents serve` mode already places the harness on customer servers, although session state in that mode is still stored by Mistral.
Developers who already use Mistral's older agent interfaces face a migration. The Agents and Conversations API, called through `client.beta.agents`, now sits under the documentation's Deprecated section with a linked guide for moving to the Agents SDK, and the durable agents feature of Workflows is flagged for deprecation. With no removal date announced, new projects should start on the Agents SDK, while existing integrations can continue to run but should be scheduled for migration. Teams evaluating the model behind the platform can find the details of its open-weight release in our report on Mistral Large 4 open weights and its modified MIT license, and a price comparison in Mistral Large 4 vs DeepSeek V4 Pro.
··········
MISTRAL STUDIO MANAGED AGENTS: CONFIRMED FACTS AND OPEN QUESTIONS.
The documented facts are specific. On November 5, 2026 Mistral added managed agents, a natural-language workflow builder, an Apps CLI, webhooks and triggers, a managed search index, default tracing with redaction, and offline and online evaluations to Mistral Studio, with all but one item in public preview and HTTP API connectors in private preview. Agents built with the `mistralai-agents` SDK can run in-process, self-hosted or Mistral-managed; persistent sessions survive crashes, default to a 24-hour lifetime and a 4-hour idle timeout, and approvals wait one hour by default. Mistral names ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services as customers.
Several points remain open. Mistral has not published prices for the new features, has not given a date for general availability, and has not said when the Mistral-hosted entries in the release notes will be available in customer VPC or on-premises deployments, despite the announcement's broader statement that the platform can run there. The maximum session lifetime after raising the default, the sandbox resource limits and the removal dates for the deprecated Agents API and durable agents are also undocumented. For companies already on Mistral's stack, the most concrete consequence of the release is the change of default: new agent work is expected to move to the Agents SDK, inside a platform whose terms are still labeled as preview.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
