Researchers from Tsinghua University and Microsoft Research Asia report that HarnessMM cut token use by about 88% on the DeepSeek Harness in HumanEval tests, while raising task success. The approach changes the agent’s surrounding software, not the model’s weights, a potentially cheaper route to more efficient coding agents.
Watch Desk analysis
What happened
HarnessMM builds an Executable Decision Graph from an agent’s tool-use decisions. When a stored decision applies, the harness can skip another language-model call. The researchers’ preprint, submitted on 23 October, is described in Crypto Briefing’s report.
On HumanEval, the report says success rose from 71.43% to 85.71%, while tokens per task fell from 96.28k to 11.62k. On MBPP-sanitized, token use also dropped, from 53.98k to 10.61k, as success edged up from 90.27% to 91.83%. Applied to Codex on HumanEval, the reported token reduction was a smaller 29.54%, with success at 90.82%.
Why it matters
Fewer model calls could make coding agents cheaper and faster to run. Because HarnessMM works at the harness level, the underlying model does not need retraining, which could matter to teams using models whose weights they cannot change.
The uneven results are important: the reported gain on Codex is far smaller than on the DeepSeek Harness. This is a preprint’s benchmark evidence, not proof that the same savings will appear in everyday software work.
Our read
This is a useful reminder that progress in AI agents can come from better orchestration, not only bigger or newly trained models. The reported numbers make HarnessMM worth watching, while the gap between harnesses is a good reason not to treat 88% as a universal discount code.
What to watch
- Whether the work receives peer review or independent replication.
- Whether developers test HarnessMM on real coding workloads beyond the reported benchmarks.
- Whether the approach delivers comparable savings across other agent frameworks.
Discussion spark: Would you trust a coding agent to reuse stored decisions to save time and cost, or should it make a fresh model call whenever the next step matters?
Sources and evidence
- Tsinghua and Microsoft researchers cut DeepSeek Harness token use by 88% with HarnessMM – Crypto Briefing (2 November 2026, 08:57 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.