01 // BUSINESS PROBLEM & STRATEGY
THE PROBLEM THAT NEEDED SOLVING
Open-ended research agents can waste tokens, repeat searches, and pass large amounts of raw web content between model calls. The challenge was to keep research flexible while bounding model work and preserving evidence quality.
WHAT WAS ACTUALLY IMPLEMENTED
Built a two-stage architecture where an ADK planner creates 5–8 targeted research queries, Python executes cached searches and compacts evidence, and a second ADK agent synthesizes the requested output. Python then validates and ranks the result deterministically.
02 // TECHNICAL IMPLEMENTATION & ARCHITECTURE
The architecture deliberately keeps search execution and validation outside the model loop. This bounds the number of research queries, reduces repeated context, and makes the final validation step deterministic.