Field note
Asynchronous agent work needs two different systems. Conductor owns the execution substrate: Claude Code, Codex, Cursor, and OpenCode can run in parallel while each task receives its own workspace, branch, files, terminal, diff, and review path. A task can mutate code and run processes without colliding with the main checkout or a neighboring job.
ChatGPT Activity addresses the other half: routing human attention. One view collects work that is running, unread, blocked, ready, or waiting for input. The operator does not keep ten tabs open just to discover that one agent needs a decision.
These layers should remain separate. A common chat surface does not provide isolation, and isolated workspaces without an attention surface create a silent backlog. A useful supervisor therefore tracks task_id, workspace_id, state, needs_human, last_transition_at, next_check_at, artifact links, and an idempotent response channel.
Conductor's current public documentation supports isolated local and cloud workspaces plus API supervision; it does not establish a microVM substrate. The durable architectural claim is the workspace contract, not an inference about the virtualization layer.