When the agent gets stuck, you sit down at its desk

Turn 41, and the agent has been fighting the same flaky test for twenty minutes. On an API-only sandbox, helping it means building a bridge: an escape hatch in your orchestrator, a way to inject commands, a way to see what it sees.

On Hoody, the agent's sandbox is a machine with URLs. You open the same terminal it's been typing into, run the test yourself, see the race condition, fix the fixture. Then you tell the agent to continue, and it continues: in a workspace that now contains your fix, because you were both in the same one.

Human-in-the-loop stops being an architecture diagram and becomes a browser tab.

You don't attach a debugger to the agent. You sit down at its desk.