A method that lets an AI coding agent reuse decisions it has already made cut the tokens DeepSeek Harness used on 98 HumanEval programming tasks by 87.93%, according to a paper posted to arXiv on 23 October. Tokens per task fell from 96.28 thousand to 11.62 thousand while the success rate rose from 71.43% to 85.71%. The four authors list affiliations at Tsinghua University and Microsoft Research Asia.
Measured per solved task, the drop is larger. DeepSeek Harness used about 134.8 thousand tokens for each task it solved before the change and about 13.6 thousand after it, roughly 9.9 times fewer, by our arithmetic (tokens per task divided by the success rate, both from the paper’s Table 1).
Business Pill · Tokens per Watt
A short explainer on tokens per watt as a measure of AI efficiency. It teaches the general idea only and says nothing about HarnessMM or this paper.
The key insight: On 98 HumanEval tasks, DeepSeek Harness used about 134.8 thousand tokens per solved task before HarnessMM and about 13.6 thousand after it, roughly 9.9 times fewer, by our arithmetic from the paper’s Table 1, while the paper reports its success rate rising from 71.43% to 85.71%.
The paper, titled HarnessMM: Recursive Self-Improvement for Harness Token Efficiency, starts from one observation. Agent harnesses “repeatedly invoke reasoning after tool feedback, even when an earlier call has already fixed the next action,” the authors write.
HarnessMM stores “exact actions, literal arguments, causal dependencies, and public guards” in what the authors call Executable Decision Graphs. When a tool result passes the stored check, the harness runs the next action without calling the model again. “New semantic decisions, unsupported evidence, and selector abstention invoke the original actor again,” the paper says.
The authors integrated the method into two existing harnesses, Codex and DeepSeek Harness, and compared each against its original version on the same tasks. On HumanEval, Codex kept its 90.82% success rate while tokens per task fell from 66.72 thousand to 47.02 thousand, a 29.54% saving, the paper says.
Model calls tell the same story. On HumanEval, DeepSeek Harness fell from 954 to 415 paid model calls, with 147 certified callback bypasses, and Codex fell from 524 to 499 calls, with 18 bypasses, according to the paper. Tokens fell faster than calls: about 9.9 thousand tokens per call before and 2.7 thousand after for DeepSeek Harness, by our arithmetic.
On 257 MBPP-sanitized tasks, DeepSeek Harness went from 53.98 thousand to 10.61 thousand tokens per task, with success rising from 90.27% to 91.83%. Codex went from 129.55 thousand to 49.22 thousand, with success at 91.83% before and 90.66% after, the paper’s Table 1 shows.
What Was Posted on X
New Tsinghua and Microsoft paper finds that DeepSeek Harness used 87.93% fewer tokens than its original version on 98 HumanEval tasks while success rose from 71.43% to 85.71%.
Agent tools like Codex and DeepSeek Harness call the LLM after each tool result to choose the next step. Often that step was already decided, like running tests once the code compiles.
HarnessMM lets the model write future commands with exact arguments, plus a check for each, such as a successful build. When a tool result passes the check, the agent runs the command and skips the LLM call.
On 98 HumanEval tasks, Codex kept its 90.82% success while using 29.54% fewer tokens than plain Codex. DeepSeek Harness rose from 71.43% to 85.71% success and used 87.93% fewer tokens than the original.
If your agent runs predictable chains like write, build, test, have the modelplan them once with pass checks instead of re-asking each step.
Rohan Paul summarised the HarnessMM paper on X on 2 November: harnesses such as Codex and DeepSeek Harness call the model after each tool result, and HarnessMM lets the model write future commands with exact arguments plus a check for each, so the agent skips the call when the check passes. His figures (87.93% fewer tokens and 71.43% to 85.71% success for DeepSeek Harness; 29.54% fewer for Codex) match the paper’s Table 1, which this piece cites directly.
Tokens per task, in thousands, for two coding-agent harnesses before and after HarnessMM, on HumanEval (98 tasks) and MBPP-sanitized (257 tasks), from Table 1 of the HarnessMM paper (arXiv 2610.32813). Pass rates shown under each row. Rounding is ours.
Where Success Fell
The paper also reports a result on Terminal-Bench. Across 18 Terminal-Bench development tasks, Codex success fell from 77.78% to 55.56% while tokens per task fell from 3,426.78 thousand to 1,089.95 thousand, the authors write in an appendix. DeepSeek Harness success fell from 55.56% to 50.00%, at 1,144.32 thousand tokens per task against 1,586.15 thousand.
The authors list the limits themselves. “Savings depend on provider input, cache and context accounting and on the availability of resolved actions,” they write, and “Authorship, selection and recovery can outweigh avoided calls.” The paper says development optimization cost is reported separately from complete-task execution.
Rohan Paul highlighted the paper on X on 2 November, summarising that the method lets the model write future commands with exact arguments plus a check for each, such as a successful build, so the agent can skip the model call when the check passes.
The Structural Read
As we read the paper, the saving comes from fewer and smaller model calls rather than a cheaper model: the authors keep the same tasks, backbone and tools, and DeepSeek Harness fell from 954 to 415 paid model calls on HumanEval, according to the paper.
As we read the results together, the method helps most where a harness spent heavily on steps that were already decided; on 18 Terminal-Bench development tasks the paper reports fewer tokens but lower success for both harnesses, and the authors say authorship, selection and recovery can outweigh avoided calls.
Yuhua Jiang, Chang Xu, Feifei Gao and Li Zhao, in the HarnessMM paper (arXiv 2610.32813)
“Agent harnesses repeatedly invoke reasoning after tool feedback, even when an earlier call has already fixed the next action.”
Three Implications
FOR TEAMS RUNNING CODING AGENTS The paper reports that Codex kept its 90.82% HumanEval success rate while tokens per task fell 29.54%, from 66.72 thousand to 47.02 thousand.
FOR BUYERS OF METERED AI PLANS DeepSeek Harness used about 2.7 thousand tokens per model call after the change against about 9.9 thousand before, by our arithmetic from the paper’s Table 1.
FOR ANYONE READING THE HEADLINE NUMBER The 87.93% saving is on 98 HumanEval tasks; across 18 Terminal-Bench development tasks the paper reports Codex success falling from 77.78% to 55.56%.
The framework says of the token: “It is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”
As we read it, HarnessMM works on the cognition side of that unit: the paper reports the same tasks solved with fewer tokens by not asking the model to re-decide settled steps. The framework is the owner’s argument and does not discuss HarnessMM.
The paper is an arXiv preprint and the results are the authors’ own measurements; we did not run the method. We did not establish which model served as the shared backbone in the HumanEval and MBPP comparisons. Neither OpenAI nor DeepSeek has commented on the results in anything we found, and Microsoft has not said whether it will use the method.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On 98 HumanEval tasks, the paper reports that DeepSeek Harness used 87.93% fewer tokens and solved more tasks once the harness reused resolved decisions instead of calling the model again, about 9.9 times fewer tokens per solved task by our arithmetic. On 18 Terminal-Bench development tasks, the paper also reports fewer tokens but lower success for both harnesses.
A note on sourcing. This piece relies on the HarnessMM paper (arXiv 2610.32813, posted 23 October 2026), read in full on arXiv on 2 November, and an X post by Rohan Paul. All benchmark figures are the authors’ own measurements. Tokens per solved task and tokens per call are our arithmetic from the paper’s Table 1. We did not run the method.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
