Approvals, Pause/Resume, and the Tasks View
Post 11 of 14
Introduction
Cowork takes actions. The control surface that keeps that safe is built into every conversation: per-action approvals with risk indicators, the ability to pause or cancel mid-task, and a Tasks view that records every run. This post is the operational user manual.
The Control Surface in Practice
Approvals. Before any sensitive action — sending an email, posting in Teams, scheduling a meeting — Cowork shows an approval prompt with:
- A preview of the content (recipients, subject, body, time, channel).
- A risk level indicator for medium- and high-risk actions.
- A Show parameters option for the technical detail.
Your choices: Action button (e.g., Send, Post, Create) — proceed once. Don’t ask again (dropdown) — proceed and skip the prompt for similar actions in this conversation only. Cancel — stop.
Pause and resume. While Cowork is working, you can:
- Pause — finish the current step, then halt. A hard pause stops immediately.
- Resume — continue from where it stopped.
- Cancel — end the task entirely.
Connection drops. If you lose your network, Cowork reconnects automatically and picks up where it left off. Progress is preserved.
Tasks view. Open Tasks in the main navigation. Three views:
- List — vertical list of conversations.
- Board — kanban by status.
- Scheduled — only visible if you have at least one scheduled prompt; lets you edit, pause, resume, or delete each.
Click any task to jump back into its conversation, inspect generated artifacts, or rerun.
Scheduled prompts. Set up by saying so in your message: “Send me a daily briefing every morning at 9 AM.” Cowork configures the schedule from your phrasing. Manage from Tasks → Scheduled or the side panel’s Schedule section.
Clarifying questions. When Cowork lacks information, it asks with a set of selectable choices. You can pick, type your own answer, or select Skip to let Cowork proceed with assumptions. The task status reads Needs user input while waiting.
Feedback channels.
- Thumbs up/down on any response.
- Document feedback on previewed files.
- Inline comments on specific message parts.
- General feedback via the menu.
Voice input. Use the microphone button in the chat input. Availability depends on your browser.
File outputs. Saved to OneDrive and SharePoint automatically. Browse them in the side panel during a conversation; access them in OneDrive afterward. Download All packages a session’s outputs as a zip.
Takeaway
The approval model is the single feature that makes agentic AI safe to use at scale. Use Don’t ask again sparingly and only inside a conversation you trust. Treat the Tasks view as your audit trail.
Start the conversation