This paper studies a common blind spot in automated coding agents: they often do far more reading and checking than a task truly needs. The authors call for task-aware execution-scope estimation — the ability to judge how hard a task is, what information it actually needs, and the shortest reliable path to a solution before spending time or money.

They introduce a formal idea called minimum-sufficient execution. That means an agent first estimates a small, plausible way to complete a task, then runs only that minimum route and expands its work only if verification fails. They also define the Agent Cognitive Redundancy Ratio (ACRR), a formal metric to quantify how much extra, unnecessary information an agent inspects. Their proposed policy is E3: Estimate, Execute, Expand.

The team tested E3 in a controlled simulator called MSE-Bench. MSE-Bench is a deterministic benchmark made up of 121 code edits in a capability-controlled environment. In this test, E3 matched the best baseline on final task success (100% success) while sharply cutting resources: it reduced cost by 85%, token usage by 91%, and the number of inspected files by 92%. E3 also outperformed a strong adaptive retrieval baseline by 16%, and these gains held up across different instruction wordings and different ways of weighting cost.

They also ran a live check with a real-model harness named LLM-Case. In that setup a gpt-4o agent edited a real open-source library and every candidate fix was graded by running the project’s pytest test suite as the true oracle. The real-system over-reading was smaller than in the simulator but still present. At comparable task success, E3 was the leanest and fastest policy. The one concrete shortfall in that run was a provider rate-limit affecting execution speed, not an incorrect edit.