20260601
This commit is contained in:
30
concepts/binding-constraint-thesis.md
Normal file
30
concepts/binding-constraint-thesis.md
Normal file
@@ -0,0 +1,30 @@
|
||||
---
|
||||
title: "Binding-Constraint Thesis(约束瓶颈论)"
|
||||
created: 2026-05-23
|
||||
updated: 2026-05-23
|
||||
type: concept
|
||||
tags: [agent, infrastructure, reliability, thesis]
|
||||
sources: [raw/papers/agent-harness-engineering-survey-2026.md]
|
||||
confidence: high
|
||||
---
|
||||
|
||||
# Binding-Constraint Thesis
|
||||
|
||||
> Agent 在真实世界中的可靠性瓶颈不在模型本身,而在包裹模型的基础设施——Agent Execution Harness。基础设施质量,而非模型能力,设定了 Agent 可靠性的天花板。
|
||||
|
||||
## 三大证据链
|
||||
|
||||
1. **工程演进证据**:从 Prompt → Context → Harness Engineering 的三阶段演进表明,约束瓶颈随工程成熟度逐步上移
|
||||
2. **跨层综合证据**:[[cost-quality-speed-trilemma]]、[[capability-control-tradeoff]]、[[harness-coupling-problem]] 三者都无法在单层内解决
|
||||
3. **开放问题证据**:五大开放问题的核心都是 harness 层面而非模型层面的问题
|
||||
|
||||
## 关键实验
|
||||
|
||||
- Bölük (2026a):仅改变 harness 格式(不改变模型),15 个 LLM 的编程能力同时提升
|
||||
- Anthropic (2026a):基础设施设置在可测量地改变 benchmark 分数
|
||||
|
||||
## 相关概念
|
||||
|
||||
- [[agent-harness-engineering]] — 总体概念
|
||||
- [[prompt-to-harness-evolution]] — 工程演进路径
|
||||
- [[agent-harness-engineering-survey]]
|
||||
Reference in New Issue
Block a user