Researchers from Tsinghua University and Microsoft Research Asia introduced HarnessMM, a method designed to optimize AI coding agents. The technique creates an Executable Decision Graph that stores previous agent decisions, allowing the system to skip redundant language model calls when predefined conditions are met.
Applied to the DeepSeek Harness on the HumanEval benchmark, HarnessMM reduced token usage from 96.28k to 11.62k per task, an approximate 88% decrease. Simultaneously, the task success rate improved from 71.43% to 85.71%. The underlying model weights remain fixed, as the optimization occurs at the harness level.
Testing on the MBPP-sanitized benchmark showed a reduction in tokens per task from 53.98k to 10.61k, with success rates rising from 90.27% to 91.83%. When applied to Codex on HumanEval, token usage dropped by 29.54% while maintaining a 90.82% success rate.
The paper, titled HarnessMM: Recursive Self-Improvement for Harness Token Efficiency, was submitted on October 23, 2026. The authors note that the efficiency gains vary depending on the specific harness architecture. The method does not require model retraining, which avoids significant computational costs.
