20 of 42

Chapter 9The harness

One brain, many bodies

The same model behaves differently depending on the harness around it, because the harness decides what it can see and do.

One brain, many bodies

Illustration

Same model, same request: "Organize my project folder and write a summary of what's in it."

What it can touch: Only what you paste or upload into the conversation.
When it asks you: Nothing to approve: it can't reach your folder.
What you get back: Advice on how to organize it, and a summary of whatever you uploaded.

Can act on your folder: no
Can open your folder: no
Model sees: Your message, Uploaded files
Runs on: The company's servers

Illustration of typical designs, not a description of any one product's current behavior.

Same model, different harness

Same model (Claude Opus 4.7), three harnesses, one research benchmark: 62.2%, 64.6% and 68.3%.¹
Same model (Claude Opus 4.5), two harnesses, a coding benchmark: 45.9% vs 55.4%, leaderboard results reported by Zhang and colleagues.²
Across ten models, "no harness performs best across all ten backends."³

When someone says "Model X scored Y," ask: in which harness?

Notes (3)

1Jiayu Wang, Weijiang Lv, Bowen Fu, Jing Fu, Jiayi Song, Lingyu Zhang, Lanxuan Xue, et al., “Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle” (preprint, arXiv, June 5, 2026), https://doi.org/10.48550/arXiv.2606.07462.
2Yunbei Zhang, Janet Wang, Yingqiang Ge, Weijie Xu, Jihun Hamm, and Chandan K. Reddy, “Stop Comparing LLM Agents Without Disclosing the Harness” (preprint, arXiv, May 7, 2026), https://doi.org/10.48550/arXiv.2605.23950.
3Yilun Yao, Xinyu Tan, Chao-Hsuan Liu, Yaoming Li, Zhengyang Wang, Zongwei Lv, Wenhan Yu, et al., “Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows” (preprint, arXiv, October 15, 2026), https://doi.org/10.48550/arXiv.2605.27922.

Chapter 9 · page 2 of 2

Turn pages with the ← and → keys.