Public preview of managed agents and apps in Mistral Studio, plus new workflow, enterprise connection, observability and evaluation capabilities. Agents defined once in the Agents SDK run in-process, self-hosted or Mistral-managed with durable background execution, approvals and tracing. Apps, workflows and agents available on Mistral Cloud today; private cloud and on-prem later in November.
Mistral Studio Managed Agents lets developers define agents with a model, instructions, tools, connectors, skills and optional execution environments, then run them in-process during development or deploy them through Studio on their own infrastructure or managed by Mistral. Deployed agents use stored sessions for durable, background execution that can recover after a crash. Developers can require human approval for selected tool calls, use per-session sandboxes, and inspect traces of model and tool activity in Studio. The Agents SDK records each step in a session, which can be interacted with live or reviewed later.
The official Managed Agents documentation identifies the Agents SDK package as the component that drives the model/tool loop and records each step in a session.
