Memory injection is 27.6% of workflow cost at depth six, and shrinking the retrieval window from 32 to 2 entries cuts it 28.7%
Total Cost of Agency decomposes multi-agent workflow spend into base prompt, inference, memory injection, miss penalty and context accumulation, and measures injected tokens with a two-pass non-billable count instead of word-count proxies — a number production observability tools do not separate out. On a 200-task enterprise benchmark against real model APIs, memory injection was about 12% of full billed cost and 13.6% of the variable cost a compile-time optimizer can touch, rising from a structural zero at depth one to 27.6% at depth six, growing linearly with depth (R²=0.9974). The lever that worked at fixed model tier: cutting retrieval window capacity from 32 entries to 2 dropped injected tokens 28.7% with accuracy change inside seed-level variation. The authors state plainly that prompt caching was not evaluated and two of five decomposition terms are zero by construction in their harness.
↳ Follow the thread