Agent AI Systems Spend Most Time Idle, Driving Up Costs Beyond Model Fees

A peer-reviewed study called 'The Harness Effect' found that model inference accounts for only about 20% of total AI agent costs, with the remaining 80% consumed by infrastructure, orchestration, tooling, and governance. Researchers found that switching orchestration layers alone — without changing the underlying model — reduced cost per task by 41%, from $0.21 to $0.12. Much of the hidden expense comes from agents sitting idle between steps: waiting on tool responses, human approvals, retries, and polling loops. CockroachDB and Stanford Digital Economy Lab research further identified that redundantly re-sent context — such as system prompts and state history — accounts for 62% of agent inference bills. Goldman Sachs projects a 24-fold increase in token consumption by 2030, meaning these inefficiencies are set to grow significantly worse over time.