Harness Engineering Enters the “Optimizable Era”: Context, State, Organization and Self-Evolution (Deep Dive)

Summary

“Agent = model + harness” is something few people dispute in 2026. The harness is all the code, configuration and execution logic outside the model [1]; it used to be seen as a mere “wrapper”, tuned by hand from experience.

But the evidence of recent months is increasingly clear: the harness affects cost, stability and even final scores no less than switching models does. More notably, a batch of work that appeared in October began treating the harness as something that can be measured, learned and automatically improved:

• At the context layer, ReFold cuts token consumption by 18%–60% with a reversible rendering layer, and AutoCompact lets the model learn for itself “when to compress”;
• At the state layer, StateHarness maintains project state explicitly, turning long-horizon consistency into a measurable metric;
• At the organization layer, Microsoft’s Agensh has 1,024 agents collaborating without a central scheduler;
• At the self-evolution layer, VERSE, RiseHarness and Self-HPO answer, respectively, “what preconditions does automatically improving a harness require”, “how can improvements be found by contrasting successes and failures” and “can the harness and the model evolve together”.

This article reviews 14 sources (5 of them background reading), breaks them down layer by layer across this four-layer structure, and offers our assessment, an outlook and practical implications for finance.

What the article covers

1. Background: why the harness has become a technology layer of its own
2. The context layer: deciding what the model “sees” at each step
3. The state layer: from “fitting it in” to “remembering what matters”
4. The organization layer: 1,024 agents without a central scheduler
5. The self-evolution layer: letting the harness optimize the harness
6. Side-by-side comparison and how the field is evolving
7. Our view
8. Outlook
9. Practical implications
10. Closing remarks

References

1. [1] Trivedy, V. (LangChain). The Anatomy of an Agent Harness. LangChain Blog, 2026-03-10 (background). https://www.langchain.com/blog/the-anatomy-of-an-agent-harness
2. [2] Weng, J. (LangChain). Context Engineering Starts with Retrieval. LangChain Blog, 2026-10-23. https://www.langchain.com/blog/context-engineering-starts-with-retrieval
3. [3] Bengre, T., Curme, C. (LangChain). Organizing Context in a Multi-Agent Harness. LangChain Blog, 2026-09-08. https://www.langchain.com/blog/organizing-context-in-a-multi-agent-harness
4. [4] Zhang, Y., Wang, J., Ge, Y., et al. (incl. Reddy, C. K.). Stop Comparing LLM Agents Without Disclosing the Harness. arXiv:2605.23950, 2026-05-07 (background). https://arxiv.org/abs/2605.23950
5. [5] Vats, N., Golev, O. The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation. arXiv:2607.22585, 2026-06-08 (background). https://arxiv.org/abs/2607.22585
6. [6] Su, Y., Tian, J., Zhang, Z., Kundu, S. ReFold: Training-Free Reversible Inter-Turn Context Folding for Long-Horizon Agents. arXiv:2610.07863, 2026-10-06. https://arxiv.org/abs/2610.07863
7. [7] Zhang, X., Zheng, L., Du, C., An, B., Dong, X. AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents. arXiv:2610.02163, 2026-10-01. https://arxiv.org/abs/2610.02163
8. [8] Zou, C., Zhou, Y., Luo, X., Zhang, X. StateHarness: Maintaining Evolving Project State for Long-Horizon Coding Agents. arXiv:2610.17616, 2026-10-13. https://arxiv.org/abs/2610.17616
9. [9] Zhan, Z., Song, T., Dong, L., et al. (incl. Wei, F.; Microsoft Research). Agensh: Scaling Organizational Intelligence to 1,024 Agents. arXiv:2609.26781, 2026-09-22; open-source repository 2026-10-24. https://arxiv.org/abs/2609.26781 ; https://github.com/microsoft/Agensh
10. [10] Lee, Y., Nair, R., Zhang, Q., et al. (incl. Khattab, O.; Finn, C.). Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052, 2026-03-30 (background). https://arxiv.org/abs/2603.28052
11. [11] Wang, Z., Ge, Y., Wang, Z., et al. (incl. Wilson, A. C.; Reddy, C. K.). VERSE: Verified Self-Evolving Optimizer for Agent Harnesses. arXiv:2610.02616, 2026-10-02. https://arxiv.org/abs/2610.02616
12. [12] Wang, R., Shi, Y., Li, Z., et al. (incl. Mi, H.). Harness Self-Evolution for Long-Horizon Terminal Agents (RiseHarness). arXiv:2610.27435, 2026-10-20. https://arxiv.org/abs/2610.27435
13. [13] Wei, Z., Yao, W., Peng, B., Wang, D., Wu, Q., Cheng, H., Meng, Y., Gao, J. Self-Evolving Coding Agents via Harness-Policy Optimization. arXiv:2610.22791, 2026-10-16. https://arxiv.org/abs/2610.22791
14. [14] Shao, Z., Wang, P., Zhu, Q., et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (origin of GRPO). arXiv:2402.03300, 2024-02-05 (background). https://arxiv.org/abs/2402.03300

HX Research · Frontier AI Deep Dive · Great Unicorn Intelligence