Coordinate agent tasks in a shared sandbox
New agent runs can plan named subtasks with dependencies and assigned files or folders. Independent tasks run
in parallel. A task waits for its prerequisites and receives their results; tasks with overlapping assigned
paths run one at a time. The Team tab and CLI show assignments and explain waiting tasks. API, Python and MCP
read the same saved view.
Direct file writes outside assigned paths are rejected. Read-only tasks cannot run shell commands. These claims coordinate cooperative agents in one sandbox; they do not isolate shell commands or symlink aliases. The parent continues to verify the combined result before completion.
Existing runs keep their original execution behavior. Individual task recovery controls and automatic subagent sizing from remaining capacity are still deferred.