Efficient Software Factory

Uber Engineering reports that some agent costs came from repeatedly checking whether a database query had finished. Each check consumed another model interaction and added intermediate output to the conversation. Its alternative puts submission, polling and result handling inside a script, returning a concise result to the model. In a five-query comparison run in one session, the first three small-result queries used 55% to 71% fewer tokens through this route. The engineering report attributes those savings to removed overhead such as schema initialization, repeated polling and step-by-step reasoning rather than to the size of the answer, so the lesson is about moving predictable coordination work into code.