Watch an agent work, and book a real call.

A working scheduling assistant. The agent's reasoning and tool calls appear in the transcript below as the conversation happens, so the loop is visible end to end.

Heads up: visitors routed through some edge regions (currently Hong Kong) may see an error from the model provider. If that happens, you can reach the same call link via How I can help.

Try it

Try the assistant.

Muhammad's assistant

Live · Claude + Cal.com

Muhammad's assistant

I'm Muhammad Arbab's scheduling assistant. Ask me to find a time, or about Muhammad, the books, or how he helps. When I use a tool or work through options, those steps appear in the transcript as they happen.

Press Enter to send. Shift + Enter inserts a new line.

Under the hood

What you're looking at.

01
Thinking rows are short reasoning passes from the model. They are streamed as the agent decides what to do next.
02
Tool calls are real function invocations the agent makes. Only two exist: check_availability and book_meeting. The agent cannot invent a time; it can only offer what Cal.com returned.
03
Tool results are whatever the tool returned. The agent reads those, then loops back to thinking. The loop is capped at five iterations so it cannot run away.
04
Final is the message the agent decided to send you. Once that arrives, the turn is done.
05
Guard rails. The agent loop is capped at five iterations per message. Requests to this endpoint are rate-limited to one every ten seconds. The model is small and fast, with a conservative token budget. These are deliberate cost and abuse controls, not afterthoughts.

This is the same loop the rest of the site writes about: decide, act, observe, decide. Showing it in the open is the point of the page.