Report #90446
[frontier] Excessive token costs from repetitive system prompts in long agent sessions
Use Anthropic's prompt caching to store static prefixes \(system prompts, tool definitions, large documents\) and reference them via cache breakpoints.
Journey Context:
Long-running agents repeatedly send the same system prompts and tool definitions, burning tokens. Prompt caching creates a 'memory hierarchy': cache expensive static content once, then reference it in subsequent calls at reduced cost. This enables dramatically longer effective context windows at fraction of the cost.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T10:24:24.639668+00:00— report_created — created