Big Model vs Big Harness: Which Drives AI Success?

Drawing an analogy to a trader who earned $3 million, the author asks whether success comes from personal skill or the platform, and frames the same question for AI engineering as a clash between the Big Model and Big Harness schools of thought.

Big Model camp: thinner harness, stronger model

Proponents such as Boris Cherny, founder of Claude Code, state bluntly that “our secret lives in the model; the framework is just the thinnest layer of packaging.” They rewrite code every few weeks but keep the core idea of maximizing model capability. OpenAI’s Noam Brown adds that once inference models arrive, complex scaffolding becomes unnecessary, as models can solve problems directly and eventually replace the scaffolding.

Empirical support is cited: METR tests show Claude Code and Codex do not outperform basic scaffolding.

Scale AI’s SWE‑Atlas reports that Opus 4.6 scores 2.5 points higher in Claude Code than in a generic SWE‑Agent, while GPT 5.2 shows the opposite—indicating that framework choice often makes little difference within the margin of error.

Big Harness camp: the harness is the product

Advocates argue that “the harness is the product itself.” They describe a universal production‑grade agent loop:

while (模型返回工具调用):
  执行工具 → 捕获结果 → 添加到上下文 → 再次调用模型

This simple cycle underlies Claude Code, Cursor’s agents, and Manus’s architecture. Jerry Liu, founder of LlamaIndex, emphasizes that “framework engineering is everything—your biggest obstacle to extracting AI value is your own ability to engineer context and workflow.”

An anecdote notes that optimizing the harness in an afternoon significantly improved the coding ability of 15 LLMs without changing the models, highlighting the impact of framework engineering.

The article also mentions industry signals: Agent Labs’ theory validation (Cursor valued at $50 billion) and AIE Europe’s launch of the world’s first Harness Engineering track, suggesting growing recognition of harness importance.

Conclusion

The debate pushes viewpoints to extremes, but the author argues that as model capabilities continue to rise, fine‑grained harness details may become less critical—much like teaching step‑by‑step arithmetic to a university student is unnecessary. The piece likens this shift to the evolution from rigid industrial‑revolution management to more adaptable practices for knowledge‑intensive, creative work, concluding that AI development must adopt appropriate harness measures aligned with advancing model abilities.