LLM agents can dramatically reduce wasted computation by estimating task complexity before acting—matching performance while cutting token usage by 91% and file inspections by 92%.
AI agents waste computational resources by re-reading information they've already seen when solving tasks. This paper introduces E3, a method that estimates task difficulty upfront, executes only the minimum necessary steps, and expands scope only when needed. On code-editing benchmarks, E3 achieves the same success rate as baseline methods while using 85% less cost and 91% fewer tokens.