New E3 Framework Cuts AI Agent Costs by 85% for Simple Tasks

This paper addresses the inefficiency of current large language model (LLM) agents, which often over-process simple tasks by re-reading unnecessary information. The authors propose that agents lack "task-aware execution-scope estimation," the ability to judge task difficulty and identify the minimal information required for completion. To quantify this, they formalize "minimum-sufficient execution" and introduce the Agent Cognitive Redundancy Ratio (ACRR). The core contribution is E3 (Estimate, Execute, Expand), a framework where an agent first estimates the task's complexity, then executes a minimum viable path, and only expands its scope if initial verification fails. Tested on MSE-Bench, a deterministic benchmark, E3 matched the strongest baseline's 100% success rate while drastically reducing costs by 85%, tokens by 91%, and inspected files by 92%. Further validation with LLM-Case, a real-model harness using GPT-4o on an open-source library, corroborated these efficiency gains, positioning E3 as a lean and fast policy for comparable task success.