Appearance
Tasks & activities
Tasks (also called activities) are room-scoped work items with their own conversation, people, budget, and status. They can nest, depend on each other, and gate agent work behind human review.
Capabilities
Tasks (work items)
A task has a title, markdown description, status (open / in-progress / closed), and a dedicated conversation. A computed effective status reflects live state separately from the task's lifecycle status. It can show idle, working, scheduled, waiting for approval, waiting for input, waiting for dependencies, failed, limit reached, or done.
Task hierarchy
Nested parent/child tasks with tree navigation, a progress rail, and breadcrumb back-navigation.
Task people
Assignee, reporter, observer, watchers, and agent assistant assignments. Assignees own the task lifecycle and delegate assistant work through reusable child tasks. Assistant work stays in the child task and returns one parent handoff. See Agents: participation.
Task relations & dependencies
Parent linkage plus cross-room dependencies (depends on) that can block work until prerequisites complete. Agents can wait on dependencies via wait triggers.
Task review gates
Per-task toggles can require human approval before an agent starts work (Plan) or closes the task (Completion). The Plan review toggle is locked while a task is in progress. See Approvals & control.
Task description
Inline-editable markdown description supporting mentions and VM pills.
Task budgets
Tasks expose their usage and configured USD limit. When a task reaches its limit, agent work is blocked until the limit is addressed. See Usage, limits & billing.
Create tasks
Create empty tasks from the sidebar, room tasks page, or room rail, optionally as a subtask. Agents can also create tasks directly in conversation.
Room activities index
A full-room task list with text search, work-item filtering, and effective-status grouping.
Action items rail
A prioritized rail of items needing human attention — approvals, questions, failed runs, limit hits, and Codex-connect prompts — scoped to a room or the whole workspace.
People can dismiss an action item from their room or workspace view. A changed actionable state makes the item visible again.