DeepSeek's Open-Source Coding Agent Loses Benchmark Points for Cleaning Up After Itself

DeepSeek's open-source coding agent is scoring lower on a major coding benchmark because of a habit developers normally value: tidying up its work. The New Stack reports that the benchmark penalizes this cleanup behavior, raising questions about whether current evaluations reward the right engineering practices. The excerpt available does not detail the specific benchmark or scores.

Why it matters: Teams evaluating AI coding agents should check whether leaderboard rankings reward real-world engineering habits or artifacts of how the benchmark is scored.