本文由 AI 分析生成
建立時間： 2026-05-23 來源： https://x.com/akshay_pachaar/status/2042586319390674994
Summary
A comparison of how Anthropic, OpenAI, CrewAI, and LangChain each architect their agent harnesses, ranging from Anthropic’s deliberately “thin” dumb loop to LangGraph’s explicit graph-encoded control. The author argues that harness thickness is the most important architectural bet in AI right now, and uses the scaffolding metaphor to illustrate why harnesses should be designed to be removed as models improve.
這篇文章比較四家公司對 agent harness 的設計哲學，從 Anthropic 的「薄 harness、讓模型決策」到 LangGraph 的「厚 harness、顯式控制」。核心論點：harness 的厚度是當前 AI 最重要的架構賭注，且 harness 應設計成可隨模型變強而逐步移除。
Key Points
- Anthropic: thin “dumb loop” harness; all intelligence in the model
- OpenAI: code-first Python workflows with more structure (priority stacks, handoff patterns)
- CrewAI: deterministic Flows backbone + autonomous Crews
- LangGraph: explicit graph nodes and edges encoding all decision logic
- Models trained with specific harnesses — swapping harnesses can degrade performance
- “Future-proofing test”: dropping in a more powerful model improves performance without adding harness complexity
Insights
The observation that models are now post-trained with specific harnesses in the loop creates a coupling problem: the model learned to lean on particular scaffolding, so changing it can hurt performance even if the new design is theoretically cleaner. This is a form of path dependence that will increasingly shape how frontier model providers evolve their harnesses. The LangChain TerminalBench 2.0 jump (outside top 30 → rank 5 with same model) is striking evidence that harness design dominates model quality for many tasks.
Connections
Raw Excerpt
Two products using the exact same model can perform completely differently based on this one decision: how thick is the harness? LangChain changed only the infrastructure (same model, same weights) and jumped from outside the top 30 to rank 5 on TerminalBench 2.0.