Who approves an AI agent's work? 22 task apps audited

Almost no task manager makes a person approve an AI agent's finished work by default. We went through the comparison data we keep on 22 task and project apps, each checked against the vendor's own pricing and help pages between October 23 and 28, 2026. 17 of the 22 run an official MCP server, so an outside agent like Claude or ChatGPT can already create, edit and complete tasks in them. 13 have no approval step for agent work at all. The other 9 have a partial one, such as a review checkpoint, an approver on one workflow step, or a preview for calendar changes. None scored above 3 out of 5 on whether agent-finished work waits for a person before it counts as done.

What did we measure?

Each app on our comparison pages is scored from 0 to 5 on eight criteria. This audit uses four of them, plus the raw cells behind them: whether you can hand a task to an AI agent, whether that agent's work waits for human approval before it counts as done, whether an agent can stop and ask you a question, and whether the app runs an MCP server that outside agents can use. A 0 on approval means the vendor documents no approval step for agent work. A 2 or 3 means there is a real mechanism, but it covers only some steps, some plans or some kinds of change. A 5 would mean every piece of agent work waits for a person by default.

Which apps have built-in agents, and which let yours in?

The 22 apps split into five groups by how an agent gets work. Nine have agents you assign tasks to inside the app: Asana's AI Teammates, ClickUp's Super Agents, Jira's Rovo and partner agents, Linear's coding agents, monday.com's agents, Notion's Custom Agents, Plane's agents, Taskade's agents and TryGoals' AI teammates. Four have an AI assistant that plans or helps rather than taking assignments: Motion, Sunsama, Reclaim and Akiflow. Five have no agent of their own but let outside agents in over an official MCP server: Todoist, TickTick, Trello, Basecamp and Any.do. One, Weft, is built around agents you bring. Three have no official path for agents at all: Things 3, Microsoft To Do and Google Tasks, where only community servers or a separate assistant reach your list.

What does approval look like where it exists?

The nine apps with a partial approval step each solve a narrower problem than signing off agent work in general. The table lists what each one documents, in the vendor's own terms.

Two patterns stand out. First, approval usually lives in the agent's own surface, not on the task: a pull request, a private draft, a preview. That works when the agent is the vendor's own, and it says nothing about an outside agent connected over MCP. Second, where a step does exist on the task, it tends to be opt-in per step or per plan, so the default is still that an agent's change counts the moment it's made.

Can an agent stop and ask you a question?

Only 2 of the 22 document it. Linear lets an agent ask for clarification inside an issue, and Weft lets an agent request input and records the decision. For the other 20, the vendor's docs don't describe a way for an agent to park a question for a person and wait, so an agent that hits an ambiguity either guesses or stops without saying why.

Why does the default matter once agents connect over MCP?

Because MCP made the outside agent the common case. 17 of the 22 apps run an official MCP server, and Claude, ChatGPT, Codex and Cursor can all connect to them. On most of those servers, an agent that calls a complete or update tool has completed or updated the task, with the same authority as the person who signed in. A per-step checkpoint inside the vendor's own agent doesn't cover that. The only place a default can sit for every agent is on the task itself: the agent proposes, a person approves.

How does Done handle it?

Done makes approval the default rather than an option. An agent working a task calls propose_completion, which posts its summary and parks the task at pending approval; it reaches done only when you approve it, unless you marked that task agent-ok beforehand. An agent can't grant itself agent-ok. When it hits a choice only you can make, it calls ask_decision, which parks the question with its recommendation and lets it move on to other work. Done is our product, so it isn't scored in the counts above.

How we built this dataset

• Source: the 22 competitor entries behind trydone.app/compare, each with its vendor sources listed beside it and a lastVerified date between October 23 and 28, 2026.
• Each cell is either a plain capability the vendor documents or a short hedge. Where a vendor's docs say nothing, the cell reads Not documented and scores as no.
• Approval scores use the comparison pages' rubric: 0 for no documented step, 2 or 3 for a real but partial mechanism, 5 only for approval of all agent work by default.
• Group counts were tallied by hand from the agent and mcp cells and checked against the totals: every app sits in exactly one group.
• Limits: vendors ship fast. ClickUp changed its plans between two checks four days apart, so treat any single cell as current only as of its date.