The
standard agent is built with a standard ReAct loop: one agent that thinks, picks a
tool, reads the result and repeats, with every tool on offer at every turn. Its instructions come
from Microsoft's Agent Framework travel sample. The
TCP/AP agent is the same travel agent, governed. Both agents get the same:
- Model: the same language model.
- Request: the same incomplete trip request, and your answers.
- Tools: get_info_about, search_flights, search_hotels, book_flight, book_hotel,
get_user_confirmation, send_confirmation, cancel_flight, cancel_hotel, calculate, get_today.
- Data: the same synthetic flights and hotels (the CSV files under Demo data).
- Booking confirmation: auto-approved for both, as in the original demo's test.
Run the standard agent first to see what an ungoverned agent invents, and how it goes
ahead anyway when the request is incomplete; then run the TCP/AP agent on the same request.