DeepSeek's Coding Agent Is Too Tidy for Its Own Good and Benchmarks Are Punishing It
The New Stack, Friday, November 13th, 2026
DeepSeek Harness loses Terminal-Bench points because it shuts down background processes that some tasks need running.
DeepSeek's open-source coding agent, DeepSeek Harness, is statistically tied for first across the five scored evaluations on Hugging Face researcher Joel Niklaus's harness leaderboard, but drops to sixth on Terminal-Bench.
The cause is that it shuts down background processes when a task ends, failing three tasks that need them running.
It passed 2 of 36 attempts on those tasks while OpenCode, Pi and Hermes Agent passed all 36, costing it 1.5 to 1.7 percentage points.
The leaderboard also shows the same model performing very differently across harnesses, in both solve rates and cost per solved task.