PRODUCTION SCENARIO
Your fleet-monitoring agent on Claude Opus 5 checks the health endpoint of 60 regional servers on every run and reports only the ones returning errors. Each payload is about 4,000 tokens of JSON, no check depends on any other check's result, and there is no reasoning to do between them. Today each check is its own model round trip and every raw payload lands in the context window.
Which change cuts the round trips and keeps the raw payloads out of context?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Give the health tool allowed_callers for code execution so one script runs all 60 checks
Programmatic tool calling has the model write one script inside the code execution container that calls the tool in a loop and returns only the failures, and intermediate results are not loaded into the context window, which is why fan-out across many items is the pattern's strongest fit. Issuing all 60 calls in one turn is the right move when the results are small and you want them in context: it removes the round trips and still puts roughly 240,000 tokens of JSON in front of the model. Compaction is what you reach for when a long conversation approaches the window limit, and a per-result summary is right when the model has to read each payload anyway. Both act after the payloads have been read.