Multi-Agent Travel Planner
A trip planner that turns a one-line request (“San Francisco to Lake Tahoe and Las Vegas, Jan 10–20, $3000, driving”) into a complete day-by-day itinerary with hotels, transport, activities, and a cost breakdown.
Architecture. Eight agents composed into a SequentialAgent pipeline: route planning → initial itinerary → refinement loop → export. The flight, hotel, and activity agents are not peers in that pipeline — they are wrapped as AgentTools and handed to a lead planner that calls them as tools, so delegation is hierarchical rather than flat. State moves between stages through shared session keys.
Self-correction. The refinement stage is a bounded LoopAgent containing a critic and a refiner. The critic validates the plan against concrete rules — no 15-hour driving days, no flights in a trip the user asked to drive, enough time to actually do the listed activities — and returns either APPROVED or three specific corrections. The refiner applies them, re-querying the flight and hotel agents as needed, and writes the revised plan back to the same session key the critic reads from. Capped at two iterations.
Grounding. Flight search runs through google_search; weather comes from the Open-Meteo API, with indoor alternatives suggested on days with a high rain probability. The final plan is converted to structured JSON and exported as a day-by-day Excel file.
Stack: Python, Google ADK, Gemini 2.5 Flash Lite, Streamlit, Pandas, OpenPyXL