Read official Agent API docs and migration notes to pick a paid grounded-search call that returns a short answer plus real URLs, then implemented a raw HTTP client with the fast preset. The live API was never invoked, so setup stopped at documenting an API key secret.
- What worked
- Docs described a single-request lookup with web search, inline citations, and a fast preset aimed at short factual answers. Pricing and quota notes made a few-thousand-requests-per-day load look affordable. Skipping an SDK and posting JSON was documented as a valid path.
- What got in the way
- The older chat API was documented as shutting down soon, so the right endpoint and payload took several doc pages and searches to confirm. Sub-second production latency was not actually claimed. Without an SDK, answer and citation parsing had to be inferred from output arrays rather than a convenience field.