@xiangxiang103: This new paper from Tsinghua × Microsoft Research Asia — my biggest takeaway after reading it: DSH is seriously underrated! The paper is about HarnessMM, a method that stops agents from "rethinking everything from scratch after each step." It uses Codex and DeepSeek Harness, two industrial-grade harnesses, as experimental subjects…

X AI KOLs Timeline Papers
agent-harness token-efficiency recursive-self-improvement decision-reuse llm-agents benchmark

Summary

The HarnessMM paper, a collaboration between Tsinghua University and Microsoft Research Asia, proposes improving agent harness token efficiency through protected decision reuse and recursive self-improvement. Experiments show DeepSeek Harness (DSH) benefits significantly: HumanEval success rate rises from 71.43% to 85.71% with 87.93% fewer tokens per problem, while Codex only saves 29.54%.

This new paper from Tsinghua × Microsoft Research Asia — my biggest takeaway after reading it: DSH is seriously underrated! The paper is about HarnessMM, a method that stops agents from "rethinking everything from scratch after each step." It uses Codex and DeepSeek Harness, two industrial-grade harnesses, as experimental subjects. In the end, Codex turned out to be just a supporting character, and DSH became the real protagonist of the whole paper. First, let's look at DSH's baseline. With the base model swapped to MiMo-v2.6-pro and nothing else added: 🔹 HumanEval: DSH 96.94% at 11.6K tokens per problem; Codex 93.88% at 62.6K tokens per problem 🔹 MBPP: DSH 97.67% at 6.2K tokens per problem; Codex 96.50% at 47.6K tokens per problem Higher scores, using only 1/5 to 1/8 of Codex's tokens. And that's DSH straight out of the box. Now look at what happens with HarnessMM installed: 🚀 HumanEval (DeepSeek-flash base): success rate 71.43% → 85.71%, tokens per problem drop from 96.3K to 11.6K — an 87.93% saving 🚀 The same modification applied to Codex: success rate stays flat, saving only 29.54% 🚀 On the MiMo base, DSH hits 97.96% with just 4.4K tokens per problem — the strongest numbers in the entire paper Why does DSH benefit so much? Look at the number of "skipped reasoning" calls: across 98 problems, DSH skipped 147 redundant reasoning invocations, while Codex only managed 18. Paid model calls dropped from 954 to 415 — more than half eliminated. The architecture is clean and easy to plug into, so optimizations from others deliver their full value. On MBPP, DSH is just as solid: success rate climbs from 90.27% to 91.83%, tokens drop from 54K to 10.6K — an 80% saving. In one sentence: the same optimization, in Codex's hands is icing on the cake; in DSH's hands it's a leap to the next level. If you're building agents, it's worth taking a serious look at DSH for your next harness. Paper link: http://arxiv.org/abs/2610.32813
Original Article
View Cached Full Text

Cached at: 11/02/26, 04:15 PM

This new paper from Tsinghua × Microsoft Research Asia — my biggest takeaway after reading it: DSH is seriously underrated!

The paper is ostensibly about HarnessMM, a method for stopping agents from “rethinking everything from scratch after every step.” It takes Codex and DeepSeek Harness, two industrial-grade harnesses, as its experimental subjects. But Codex ends up playing a supporting role, and DSH becomes the protagonist of the whole paper.

First, let’s look at DSH’s own fundamentals. Swap the base model to MiMo-v2.6-pro, add nothing else:

🔹 HumanEval: DSH 96.94%, 11.6k tokens per problem; Codex 93.88%, 62.6k tokens per problem 🔹 MBPP: DSH 97.67%, 6200 tokens per problem; Codex 96.50%, 47.6k tokens per problem

Higher scores, using only 1/5 to 1/8 of the tokens Codex spends.

And this is DSH exactly as shipped from the factory.

Now look at it with HarnessMM installed:

🚀 HumanEval (DeepSeek-flash base): success rate 71.43% → 85.71%, tokens per problem drop from 96.3k to 11.6k, saving 87.93% 🚀 The same change installed into Codex: success rate stays flat, saving only 29.54% 🚀 Under the MiMo base, DSH hits 97.96% with just 4400 tokens per problem — the strongest set of numbers in the entire paper

Why does DSH capture so much more of the benefit? Look at how often reasoning gets skipped: across 98 problems, DSH skipped 147 redundant reasoning calls, versus only 18 for Codex. Paid model calls were cut from 954 to 415 — more than half simply gone. The architecture is clean and easy to plug into; optimizations from others pay off in full when dropped in.

On MBPP, DSH is just as steady: success rate rises from 90.27% to 91.83%, tokens drop from 54k to 10.6k, an 80% saving.

In one sentence: the same optimization, in Codex’s hands is icing on the cake; in DSH’s hands it’s a leap to the top.

For anyone building agents, take a serious look at DSH as your next harness.

Paper link: http://arxiv.org/abs/2610.32813

Similar Articles

@katcat_AI: https://x.com/katcat_AI/status/2116724336128381145

X AI KOLs Timeline

This article provides a technical breakdown of dsh-superagent-prototype, open-sourced by the DeepSeek Harness team, analyzing how a programming agent capable of continuously taking on orders handles authorization, task distribution, Worker scheduling, and interruption recovery based on transactions and persistent queues. The article emphasizes that this prototype is for research reference only, and is not an official DeepSeek product.

Harness-Delta Attribution: Dissecting Gains From Harness Evolution

arXiv cs.AI

提出了Harness-Delta Attribution（HDA）评估协议，将agent harness演化带来的性能增益分解为过拟合、测试时计算扩展和可泛化改进三部分，发现11个executor模型和3个框架中过拟合与测试时扩展平均解释了54%和14%的增益，搜索集收益往往无法迁移到held-out任务。

@Xudong07452910: This latest paper, Scaling Laws for Agent Harnesses, is a must-read for those working on Agent Harnesses. It highlights a key point: Agents don't necessarily become stronger by running more tokens, tuning more tools, or looping more rounds. What really matters is that these...

X AI KOLs Timeline

This paper proposes Effective Feedback Compute (EFC) as a scaling coordinate for measuring Agent Harness performance, emphasizing that effective feedback is more important than raw compute, with important implications for Agent system design.