When an agent stops

Every turn an agent takes ends one of three ways: it delivered something (a commit landed), it asked a question, or it did nothing — the CLI it was running on was signed out, crashed, or ran out of quota. That third outcome is what lands an agent in the board’s Stuck column, even when there’s no question on screen to answer.

Answering a stuck agent

Open the agent’s page and you’ll find a field to answer it, right on the card it stopped on. Whatever you type becomes its next prompt — the same mechanism whether it asked something specific or simply went quiet. There’s nothing else to configure; typing and sending is the whole act.

A CLI limit

The machines a project runs on share one login per CLI, and that login can hit a limit. When it does, every agent currently on that CLI is paused together, and the board says why in the CLI’s own words next to the affected agents. A paused agent offers Run on Claude instead — pressing it moves that one agent to Claude Code, and it picks up on the machine’s next poll.

Stopping an agent

Stopping an agent is different from it getting stuck — it’s something you choose. Pressing Stop discards the agent’s current work outright: its cards go back to Open, and any card it had already worked wears a small “stopped” stamp with how long the agent spent on it before it was stopped. The stamp never says why it was stopped or what it cost — just that it happened, and when.

See Answering a question for where and how to answer, and A run, turn by turn for reading everything that led up to a stop.