Towards a responsive agentic behavior
A harness where the agent keeps talking to you while its tools are still running. Tool calls do not block the conversation: they become tasks the agent can look in on, report about, and be asked to abandon.
Everything here is canned. No model is called, so the same thing happens every time and nobody's bill moves. What is real is the harness: the same board, the same background tasks and the same widgets the live version uses. The panel on the right shows the board the model would have been handed on each turn.
Use Play to watch it in real time, Pause to stop the clock for everything at once, and Step to advance one beat at a time.
Scenarios
One tool call, then one that wedges
The simplest shape first: a question, a tool that takes a while, and the answer arriving on its own. Then a follow-up whose tool reports progress and then stops dead - the harness notices it is overdue, and the agent tails it, kills it and starts it again.
Run itThree at once, with a conversation over the top
Three lookups start together and the conversation carries on regardless: a joke while they run, a progress check that makes the agent tail every task, and then the answer arriving on its own once the last one settles.
Run itA second question, about something else entirely
A weather lookup is still running when the user asks an unrelated scheduling question, so two different tools are in flight at once. The calendars come back first and are answered on the spot; the weather lands afterwards and has to say which question it belongs to.
Run it