DeepSeek’s coding agent is too tidy for its own good — and benchmarks are punishing it

A Hugging Face researcher's coding-agent leaderboard found DeepSeek Harness statistically tied for first across most evaluations, yet it falls to sixth on Terminal-Bench. The cause is a built-in habit of stopping background processes when a task ends, which breaks three tasks that depend on them to keep running. In the researcher's tests, DeepSeek Harness passed only two of 36 attempts on certain tasks, while several competing agents passed all of them. That single behavior accounted for more than the entire gap between DeepSeek Harness and the leader on one model. The leaderboard's changelog confirms the behavior belongs to the harness itself, and the results were left unchanged. The broader study also showed that holding the model constant still produces large swings in success rate and cost depending on the harness used. The takeaway is that the agent wrapper deserves as much testing as the underlying model.