A new preprint shows AI coding agents can skip predictable model calls, using far fewer tokens while succeeding more often
AI coding agents talk to themselves a lot. A new paper from Tsinghua University and Microsoft Research Asia suggests much of that chatter is unnecessary.
Applied to the DeepSeek Harness, the researchers’ technique, called HarnessMM, cut tokens per task from 96.28k to 11.62k, an approximately 88% reduction. Task success rates went up at the same time, which is not usually how efficiency trade-offs work.
What HarnessMM actually does #
The preprint is titled “HarnessMM: Recursive Self-Improvement for Harness Token Efficiency.” It was submitted on October 23, 2026, by Yuhua Jiang, Chang Xu, Feifei Gao, and Li Zhao.
Some vocabulary first. A “harness” is the scaffolding around a language model that turns it into an agent. It decides when the model gets called, which tools it can use, and how results get fed back into the loop.
Tokens are the chunks of text a model reads and writes. Every token costs compute, so an agent that burns fewer of them is cheaper and faster to run.
HarnessMM gives an AI agent something like that muscle memory. The method builds what the authors call an Executable Decision Graph, a structure that maps out the agent’s decision-making after it receives feedback from a tool.
Inside that graph sit “guards.” These store decisions the agent has already worked out. When predefined conditions are met, the harness uses the stored answer and skips a fresh call to the language model entirely.
The “recursive self-improvement” part refers to the harness configuration itself getting better over time. Crucially, the actor model’s weights stay fixed. Nobody retrains the underlying model. All the gains come from smarter plumbing.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The numbers #
The headline result comes from the DeepSeek Harness running on the HumanEval coding benchmark. With HarnessMM, task success climbed from 71.43% to 85.71%.
Tokens per task, meanwhile, fell from 96.28k to 11.62k. The agent got right answers more often while using roughly an eighth of the text budget.
The researchers also tested the method on MBPP-sanitized, a second set of coding tasks. There, success edged up from 90.27% to 91.83%, while tokens per task dropped from 53.98k to 10.61k.
HarnessMM was also applied to Codex on HumanEval. The results were more modest but still meaningful: success held at a high 90.82%, and token usage fell by 29.54%.
That gap is worth noticing. The DeepSeek Harness saw a dramatic cut, while Codex saw a solid but smaller one. The amount of waste available to trim seems to vary depending on the harness the method is applied to.
Why the DeepSeek Harness matters here #
The DeepSeek Harness is a fairly fresh project. It was open-sourced on August 13, 2026, and has since piled up tens of thousands of GitHub stars.
For teams building AI coding tools, the most important detail is that no retraining is required. Retraining a model is expensive and often impossible when you’re using someone else’s weights. A harness-level fix sidesteps that bottleneck entirely. That makes the approach potentially portable. The Codex result hints that the method isn’t locked to a single framework, even if the size of the payoff differs from one system to another.
There are reasons for caution, though. This is a preprint, which means it hasn’t yet been through formal peer review. HumanEval and MBPP-sanitized are established benchmarks, but benchmarks are tidy by design.
The guard mechanism also raises an obvious question. Skipping model calls works beautifully when the next step really is predictable. The paper’s own results suggest that risk was manageable on these tests, since success rates rose rather than fell.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.