Agents are good at doing work and bad at knowing which calls are yours to make. Left alone they guess, and you find the guess three days later in a migration you did not want. Loopyard gives them a better option: park the decision and carry on with everything else.
When an agent hits a real fork, it does not bury the question in a paragraph of prose. It posts a card: the question, the options with their tradeoffs, a recommendation when it has one, and a skip. You tap and it continues. You ignore it and the decision waits.
That waiting is the part that matters. A decision here is stored state, not a notification that disappears. It survives server restarts, agent restarts, and you closing your laptop. Come back Monday and it is still there, still answerable, and your answer still reaches the agent that asked.
Answer from wherever you are. The card sits inline in the chat. The operator screen lists everything blocked on a human across every workspace. When a few pile up, the Reviewer deals them out one at a time so you can clear the backlog in a minute of tapping. Replying in plain chat works too, and your message becomes the answer.
Bigger calls ride the same rails: forking a workspace, merging a branch back, deleting a workspace. The agent proposes, you approve. When it needs an API key it asks through a masked field, so the key never appears in the conversation.
