What happened
Mistral’s Managed Agents documentation says the product and the Agents SDK, mistralai-agents, are in public preview. The company cautions that the API, behavior, and limits can change during preview, and updates may require code changes. It describes Managed Agents as a way to run autonomous agents for hours or days, resume after a crash, and run many sessions in parallel, while retaining control over what an agent can access and do.
A developer defines an agent’s model and instructions, tools, connectors, skills, environment, and which tool calls require human approval. The SDK drives the model-and-tool loop and records each step in a session. Mistral frames production operation as requiring durable execution, runtime infrastructure and per-session environments where shell or file access is needed, protected secrets, and traces for later auditing. Managed Agents is intended to provide those elements.
The documentation distinguishes local development from deployment. A builder can run the agent in their own process while developing, then deploy through Studio either self-hosted on their servers with agents serve or Mistral-managed with agents deploy. The documentation’s conceptual model separates the model, which chooses whether to answer or call a tool; the harness, which runs the loop and tools; the session, which records messages, calls, and results; and the environment, the computer or sandbox where work happens. Depending on deployment, the harness and tools run in the developer’s process or on servers, while a sandbox may run locally, on Mistral, or in a hosted environment.
Mistral recommends Chat Completions for a single model request without an agent loop, Managed Agents for autonomous loops, and Workflows for processes with explicit steps, retries, and branching. Workflows are generally available; Managed Agents and Apps are in public preview. The supplied documentation excerpt gives no price or specific numerical limits. A separate supplied summary says private-cloud and on-premises deployments will follow at the end of November, but the documentation excerpt does not substantiate that timing.
What the archive adds
Build Enterprise Agents, Workflows & Apps in Mistral Studio covered Mistral Studio’s managed agents, apps, workflows, and enterprise features. This is a follow-up availability clarification, not an unrelated new product thread: the current source makes the preview status of Managed Agents and GA status of Workflows explicit.
Trend signal
This supports harness architecture outlives model choice: Mistral’s documentation treats the harness, tools, session, and environment as separate operational components around a model. The documentation describes a platform architecture, however, not evidence that this boundary is winning adoption or producing better outcomes.
Why now
The source presents a deployment-state distinction that matters to teams deciding whether to test or rely on the service: Workflows are generally available, while Managed Agents and the SDK remain preview software subject to changes in API, behavior, and limits. That is a practical planning signal, not a stated launch date.
Why it matters
For builders, the useful choice is not simply “agent or no agent.” A single request, a process with explicit steps, and an autonomous loop have different operating requirements. Mistral’s framing gives teams a way to map those requirements: use Workflows when retries and branching need to be explicit; consider Managed Agents when tasks must continue over time, survive crashes, or run in parallel.
The deployment choice also makes operational ownership visible. Self-hosting may preserve control over where the harness and tools execute, while a managed deployment shifts that execution to Mistral. In either case, builders need to decide what the agent can reach, which actions require human approval, where its environment resides, and how traces and secrets are handled. The preview caveat is material: APIs and limits can change, so production plans should account for migration work rather than assume stability. Earlier coverage: Build Enterprise Agents, Workflows & Apps in Mistral Studio.
Limits and uncertainty
The supplied source is product documentation, not an independent evaluation. It describes capabilities and intended deployment modes but does not establish reliability, task-completion rates, security outcomes, performance, or comparative cost. It gives no pricing, concrete concurrency or duration limits, or detailed service-level commitments. Public preview means the documented behavior may change. The source also does not establish which regions or customer plans can access the features. The private-cloud and on-premises timing appears only in a supplied summary, not in the excerpted primary documentation, so it should not be treated here as independently confirmed.
Skill to build now
Before testing a managed-agent platform, write a short deployment contract for one real task: permitted tools and data, actions requiring human approval, expected duration, recovery behavior, session record requirements, and the environment where shell or file access can run. Map each item to the source’s model, harness, session, and environment concepts. This makes a preview evaluation useful without mistaking product availability for production readiness.
Reader takeaway
Mistral draws a clear line between explicitly stepped Workflows, now generally available, and autonomous Managed Agents, still in public preview. For a real deployment decision, assess the loop’s duration and recovery needs alongside tool permissions, environment, and operational ownership—and budget for change while the agent stack remains preview software.