Researchers changed how DeepSeek Harness handled decisions and context, cutting model-token use by 87.93% per task in a coding test while raising its success rate on the same problems.
The Oct. 23 preprint measured 98 held-out HumanEval tasks, rather than the benchmark's full 164-task release. The modified agent solved 85.71% of them, compared with 71.43% for the baseline. Average input use dropped from 92,450 to 7,090 tokens per task, outweighing a rise in output use from 3,830 to 4,530. Total use averaged 11,620 tokens per task after the changes, compared with 96,280 before them.
HarnessMM, the paper's action-reuse method, saves a decision the model has already made along with conditions for applying it after a tool returns. The agent can carry out that decision if the result satisfies those conditions. Otherwise, it asks the model to assess the new information.
The measured savings reflect more than that mechanism. Researchers also revised context management and how the agent ended tasks, while keeping the model, tools, tasks and scoring method fixed within each baseline–modified pair. Paid model calls declined from 954 to 415, but only 147 instances were certified as calls avoided through action reuse.
The second test covered 257 problems in a held-out MBPP-sanitized subset. DeepSeek Harness averaged 10,610 tokens per task, versus 53,980 for the baseline, and its success rate increased from 90.27% to 91.83%. For Codex, lower token consumption on that test came with a decline in completion: 90.66%, compared with 91.83% before modification.
Codex had a different result on HumanEval: its success rate remained 90.82% as average token use fell from 66,720 to 47,020 per task. The HumanEval and MBPP-sanitized results cover different sets of problems.
This experiment concerns decisions within individual coding tasks. DeepSeek Harness also runs a continuously operating internal system for team coordination, described in earlier EconoTimes coverage.
The results measure tokens, not dollar savings, completion time or energy use. The preprint says the opportunity to avoid model calls depends on how often a saved decision fits later tool feedback; provider accounting affects what lower token use means for charges.
