Telemetry is what makes an agent improvable
Technology
You cannot tune what you cannot measure. Much of what we changed in our nine agent Gemini pipeline came from watching it run in production, not from rereading the prompt.
BigQuery is where that telemetry lands. We also use it as a plain serverless warehouse when a site needs queryable data behind it, which is a smaller job than the product is sold for and still the right tool.
Where
Where BigQuery earns its place
- Telemetry that outlives the request
Our nine agent pipeline writes structured telemetry to BigQuery alongside Cloud Logging. A log tells you what one request did. A table lets you ask what a thousand requests have in common, which is the question you have when you are choosing what to optimize.
- A warehouse can back a marketing site
On Cottrell Trailers we used BigQuery to hold location data and site behavior, then queried it for dynamic filtering and a Google Maps integration. Not a data platform engagement. Just the store that let the location finder answer questions quickly.
- You are billed for what you scan
The cost model punishes wide select statements and unpartitioned tables. For small, hot, transactional reads it is the wrong shape entirely, and Postgres or Firestore will be cheaper and faster. We reach for it when the query is analytical and the table is large.
