User Actions: when the AI agent hands over to you

An AI agent cannot do everything: it cannot open a Stripe account that requires identity verification, generate an API key behind two-factor authentication, or draw a game’s sprites. Instead of failing or inventing a workaround, it creates a User Action: a user-action ticket that states exactly what you need to do.

The handover is explicit: the AI works, you handle the step that is yours, then the AI resumes. Enable the User Actions plugin from the Marketplace; it has no UI of its own and plugs into the Mission Control board.

Actions land in a dedicated column right after Todo, with a counter. Their cards pulse to draw attention and cannot be dragged. The column disappears when it is empty.

Each card shows its parent ticket — the agent that asked for it — and the blocks F-20 chip shows which work is waiting. An action whose prerequisite is not yet done or integrated stays in Pending instead of joining the column.

Open an action to read its instructions. Singularity’s prompts enforce a format: what is already done, why you, numbered steps with the exact accounts, menus and fields, then the observable done criterion. A secret is never requested in the ticket: the agent says where to enter it, for example in the project secrets.

The detail view also shows the project, the parent ticket and the dependencies. Steps and criteria are not structured fields: everything lives in the instructions text.

The agent creates the action with the user_action_create MCP tool: title, instructions, priority from low to critical, and a blocking option. The action is attached to the agent’s ticket and awaits your approval.

With blocking: true, the agent’s ticket depends on the action: its card shows blocked by F-24 and a 0/1 done bar. The agent then calls orchestration_wait_for_ticket and its run pauses without consuming quota. The user_action_list tool lets it find the workspace’s actions. See dependencies.

You can also record a human task yourself. In New ticket, choose the Human agent: the ticket is created as a user-action, with no model or reasoning level, and is never sent to an AI. It joins the column like the agents’ actions.

The board search also filters this column, by title, description or number.

Type a few letters in the Mission Control search bar: only matching cards remain visible, including in the User action column. Clear the search to get the full board back.

Once the step is done, click the green Mark as done button under the card. The action moves to done and the agent’s prerequisite is met: the bar switches to 1/1 done and, when all its dependencies are done, the scheduler restarts the agent’s ticket, which returns to In progress.

An action without an agent can also be handed to an AI with Set agent…, and replying to the ticket moves it to in progress and starts the assigned agent. In the Horizon iPhone app, the column and creating a “Human” action are available; the done button is not yet.