Mistral’s 5 November 2026 release notes introduce Managed Agents and the Apps CLI as public previews. The release has mixed availability: HTTP connectors are private preview, while displaying Studio apps inside Vibe Code Web is generally available. This is an expansion of Studio, not a new platform launch; the workflow visualizer was already listed on 14 October.
Separate the agent, the process and the interface
The Studio announcement connects three layers: agents investigate and act, workflows coordinate steps and approvals, and apps give people an interface. Mistral promotes flexible cloud and private deployments. For an operator, the useful question is which components run where in the specific deployment being purchased or tested.
Check where state lives
In the documented self-hosted mode, your servers run the agent loop and tools, while Mistral stores sessions. Deployed runs can resume after crashes from the last completed model or tool call. Developers select which tool calls require approval. This public preview SDK may change APIs, behavior or limits, requiring code updates.
Check entitlement before planning an app rollout
Studio apps package components such as an interface, API, database, agent and workflow into one project. Studio supplies hosting, authentication and access controls. The current guide says Free plans, and Pro or Team without pay-as-you-go, cannot create deployed apps even when deployment commands are visible. Check the organization’s actual entitlement rather than assuming a working local prototype can be published.
Our proposed acceptance checks
- Choose one bounded task and synthetic or approved non-sensitive inputs. Write down what success and an unacceptable action look like before running it.
- Exercise an interrupted run, a denied approval and a duplicate incoming event. Check external systems for repeated changes, not only the agent’s final answer.
- Record who can inspect stored work, how long it remains available, and who owns recovery. Confirm those answers for the selected service and contract.
- Measure completion quality, latency, model usage and human review effort together. Set a spending limit and a rollback decision before expanding access.
These are proposed checks, not reported test results. We have not deployed the SDK or measured reliability, savings or security. Preview availability is a reason for a controlled evaluation, with production acceptance decided from evidence.