harness-scalinglisted

> **Attribution:** Derived from [arXiv:2605.26112v1](https://arxiv.org/abs/2605.26112v1) — *From Model Scaling to System Scaling: Scaling the Harness in Agentic AI* by Shangding Gu (UC Berkeley), May 2026. Reference implementation: [CheetahClaws](https://github.com/SafeRL-Lab/cheetahclaws). # Harness Scaling for Agentic AI Agent performance emerges from the **interaction among multiple components**, not from model capability alone. This skill encodes the insight that investing only in stronger foundation models while neglecting the surrounding harness — memory, context assembly, skill routing, and governance — yields unreliable agents regardless of model strength. ## When to Activate Activate this skill when: - Designing a new agentic system from scratch and choosing architectural components - Debugging persistent agent failures that survive model upgrades - Evaluating whether to invest in a stronger model vs. improving harness design - Auditing a production agent for reliability, auditability, or safety gaps - Implementing multi-agent coordination systems - Choosing memory, context, or routing strategies for an agent deployment - Answering: "why does my agent confidently do the wrong thing?" ## Core Concepts ### The Harness Equation Agent harness performance is a function of six interacting components: ``` 𝒫_H = Φ(ℛ, ℳ, 𝒞, 𝒮, 𝒪, 𝒢) ``` | Symbol | Component | Role | |--------|-----------|------| | **ℛ** | Reasoning Substrate | The foundation model; improved via