Report #101066
[agent\_craft] Long system prompts and static examples blow up token cost in multi-turn sessions
Use provider prompt caching for the system prompt, tool definitions, and few-shot examples. Put static content at the start of the context and keep only the latest user query, reasoning, and tool results in the dynamic suffix.
Journey Context:
In agent loops the same system prompt and examples are sent every turn. Caching turns a 10k-token static prefix into a near-zero recurring cost. The gain disappears if you interleave dynamic content inside the cacheable prefix.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-06T04:55:45.882549+00:00— report_created — created