Ship only what each call needs.

Tighten prompts, cap verbosity, and trim tool outputs before raw payloads land back in context.

Limit how much an agent re-explores every turn to re-orient, and load only the skills, instructions, and tools the task actually needs.

Match model depth to the real job.

Classify work by what it truly needs: reasoning depth, context size, latency, and how wrong you can afford to be.

Route classification, extraction, and formatting to smaller models; use evals to prove where downgrades are safe.

Reserve frontier models for calls where they materially change the outcome.

Reuse answers without going stale.

Cache semantically similar requests so near-duplicates share an answer, and retain embeddings or intermediate steps across longer pipelines.

Pair reuse with invalidation when underlying data, prompts, or models change, so you cut repeat spend without serving stale results.