Microsoft Conductor moves agent execution into container profiles Microsoft Conductor moves agent execution into container profiles

Moving more than a shell command

Conductor defines multi-agent workflows in YAML, with routes selected through expressions rather than a model deciding the next step. Its existing workflow tools include parallel agents, reusable sub-workflows, shell scripts and direct MCP calls. Human gates pause a workflow for a decision. That orchestration layer is separate from the agent doing the work: a predictable route through a workflow does not make the model’s individual answers predictable. The release changes where those agent steps execute, while preserving the distinction between workflow control and model behavior. Validation catches undeclared dependencies and stale references before a run, while iteration and timeout limits bound how long a workflow continues. The command-line tool also exposes a web dashboard and a Fleet Manager, giving operators workflow state and human decisions outside the agent’s own conversation.

Previously, Docker execution profiles covered script steps. The new agent placement also moves the model interaction loop and its stdio MCP child processes into the selected environment. Docker can stage skills and plugins with the workspace; Azure Container Apps cannot stage those dependencies and requires them baked into the runner image. Moving a workflow between these environments therefore involves its tools and dependencies, not only changing a model name.

Recovery preserves files, not a suspended turn

The implementation review identified a specific recovery problem: an agent-only Docker run could resume by staging files over a workspace the agent had already changed. The merged fix records agent dispatch and refuses attachment when the staging marker is missing. The intact-marker case reconnects without copying the bundle again. This matters for workflows that generate files before failing, because preserving those files is different from restarting the agent. The documented resume model reruns remote agents from their initial inputs; it does not restore their suspended Python execution. Runner image names are recorded, but resolved image digests are not pinned across resumption.

The release also adds targeted interruption of an active call and delivers agent-scoped secrets to stdio MCP child processes. Remote profiles default to excluding the control process environment. Docker checks for protocol version 2 before dispatch, rejecting an outdated runner image early. Each concurrent invocation gets its own runner; sequential calls can reuse an idle one. These mechanics connect execution placement to capacity planning and image management. A workflow with parallel agent steps needs enough execution capacity for those calls, even when its routing configuration remains unchanged.