本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/akshay_pachaar/status/2042586319390674994
Summary
Akshay Pachaar maps four companies’ bets on harness thickness: Anthropic (thin dumb loop, trust the model), OpenAI (slightly thicker, code-first with explicit handoffs), CrewAI (deterministic backbone for routing + autonomous crews), LangGraph (explicit control graph, logic in the harness). The scaffolding metaphor frames this as temporary infrastructure designed to be removed as models improve. The co-evolution problem: models trained with specific harnesses may perform worse if the harness changes.
Akshay 比較了四家公司對 harness 厚薄的不同押注。Anthropic 選擇最薄的 dumb loop，相信模型自身會越來越強；LangGraph 在另一端，將所有決策邏輯編碼到圖結構中。關鍵洞見是：scaffolding 就像建築鷹架，隨著建築完成應逐步拆除，但模型已經學會依賴它，所以不能貿然移除。
Key Points
- Anthropic: dumb loop, model makes all decisions, harness manages turns only
- OpenAI: code-first Python, priority stacks, explicit handoff patterns
- CrewAI: deterministic Flows layer + autonomous Crews
- LangGraph: explicit node/edge graph, every decision is hardcoded
- Co-evolution trap: models trained on specific harnesses break when harness changes
- Future-proofing test: does a better model improve performance without adding harness complexity?
Insights
The LangChain harness-only upgrade (same model, same weights) jumping 25+ positions on TerminalBench is striking — it makes the harness the product, not the model. The co-evolution problem is underappreciated: if models are post-trained against a specific harness, you can’t swap the harness without a corresponding model update. This creates lock-in between model checkpoints and infrastructure.
Connections
Raw Excerpt
“Build scaffolding that’s designed to be removed. But remove it carefully, because the model learned to lean on it.”