Mistral opens managed agents with three execution modes Mistral opens managed agents with three execution modes

Execution becomes a deployment choice

The Agents SDK separates an agent definition from where it runs. Developers can execute it inside their application, serve a worker themselves, or deploy it to Mistral. For deployed agents, the release notes describe persistent sessions that continue after an application exits and recover from the last completed step after a crash. A tool call can wait for approval, then continue later. This changes the unit a developer operates: a stored task with a lifecycle, rather than a single model response. Studio also records model and tool calls as traces, making individual turns inspectable. Local development remains a separate execution choice.

The platform connects actions to applications

The broader Studio announcement connects agents, workflows and employee-facing applications. Agents can use skills, code sandboxes and delegated subagents, while workflows coordinate fixed steps and human decisions. Applications provide the interface through which staff use those processes. Mistral supplies hosting, authentication and access controls, and describes deployments across its own infrastructure, customer clouds and on-premises environments. That brings several operational responsibilities into the platform instead of leaving every application team to assemble them separately.

Permissions predate the new runtime

The access controls underneath this release have their own history. On June 24, Mistral introduced organization and workspace controls that can disable individual connector tools. An administrator can leave a connection available while blocking its write operations. Scoped keys and service accounts distinguish automated work from a developer’s personal identity. Multi-account connectors also separate logins to the same service. These controls matter when an agent continues running after its initiating user leaves: connectivity alone does not establish which identity should act or which operations it should perform. The June release also introduced a connector debugger that checks eleven stages of establishing an MCP connection.

Instructions have a deployment history too

Mistral added versioned prompts and skills to Studio on July 9. Each version is fixed, attributed to an owner and recorded in a change history. Teams can compare versions, roll back a change and promote assets through staging and production labels. The company describes links between operational traces and the instruction version behind an output. This supplies a second record alongside the new execution history: what the agent was told, as well as which tools it called. For an investigation into unexpected behavior, both records matter. A successful tool call can still carry out the wrong instruction, and an instruction change can affect many sessions without changing application code.