Researchers propose E3 (Estimate, Execute, Expand), a framework for task-aware execution-scope estimation that allows AI agents to judge task difficulty and information needs before committing budget. The method formalizes minimum-sufficient execution and the Agent Cognitive Redundancy Ratio (ACRR) to prevent agents from over-reading files and dependencies.
On the MSE-Bench benchmark of 121 edits, E3 matches the strongest baseline's 100% success rate while cutting cost by 85%, tokens by 91%, and inspected files by 92%. It also beats a strong adaptive retrieval baseline by 16%, with gains surviving held-out instruction wording and various cost weightings. A companion harness using gpt-4o on a real open-source library confirmed E3 as the leanest and fastest policy at comparable task success.
The authors frame this as a step toward engineering-grounded AI (EGAI), where agent effort is anchored in the engineering reality of the task rather than maximum-context strategies. The framework and benchmark are released to support this shift.