The Harness Effect: orchestration overhead drives 80% of agent costs, not inference

Short summary

A peer-reviewed study called 'The Harness Effect' found that model inference is only ~20% of total agent cost; infrastructure, orchestration, tooling, and governance consume the other 80%. Switching orchestration layers cut cost per task by 41% — more than switching models. Re-sent context alone accounts for 62% of agent inference bills, and Goldman Sachs projects a 24x increase in token consumption by 2030. The article breaks down four categories of agent waiting costs: tool latency, human approval gates, retry backoff, and polling loops.

• •Inference is only ~20% of agent cost; orchestration and infrastructure eat 80%
• •Switching orchestration layers reduced cost per task by 41%, more than switching models
• •Re-sent context (system prompts, tool definitions, state history) accounts for 62% of inference bills

Generated with AI, which can make mistakes.