A new July 14 paper argues that many coding agents do too much work before they do the right work.
The arXiv paper, “Do AI Agents Know When a Task Is Simple?”, proposes E3: Estimate, Execute, Expand. The idea is straightforward. An agent should estimate the smallest reliable scope for a task, execute that minimum path, and expand only if verification fails.
The authors frame the missing skill as task-aware execution-scope estimation. Instead of starting with a maximum-context-first strategy, the agent asks how difficult the task is, what information it truly needs, and what the shortest reliable path looks like.
The reported savings are large
On MSE-Bench, a deterministic benchmark of 121 edits in a controlled simulator, the authors report that E3 matched the strongest baseline’s 100% success rate while cutting cost by 85%, tokens by 91%, and inspected files by 92%.
The paper also describes a real-model harness called LLM-Case, where a live gpt-4o agent edits a real open-source library and patches are graded by running the project’s pytest suite against an oracle. The authors say the over-reading effect is milder there but still present, with E3 remaining the leanest and fastest policy at comparable task success.
Those numbers are author-reported. They should not be read as a benchmark of deployed commercial agents. The useful point is more general: agent quality is not only about final correctness. It is also about how much context, cost, time, and tool work the agent burns on the way there.
Minimum sufficient execution is a product problem
Many agent interfaces reward exhaustive behavior. The agent opens more files, gathers more context, writes a longer plan, and sounds safer. Sometimes that is exactly right. For ambiguous migrations, security-sensitive changes, or unknown codebases, the agent should broaden its view.
But simple tasks are different. A one-line fix does not always need a small codebase audit. Over-reading can slow the user down, increase token spend, hit rate limits, and expose more context than the task needs.
E3’s framing is valuable because it makes scope a first-class behavior. The agent starts small, verifies, and then earns additional exploration when the first path fails.
