20260601
This commit is contained in:
34
concepts/adaptive-harness-simplification.md
Normal file
34
concepts/adaptive-harness-simplification.md
Normal file
@@ -0,0 +1,34 @@
|
||||
---
|
||||
title: "Adaptive Harness Simplification(自适应 Harness 简化)"
|
||||
created: 2026-05-23
|
||||
updated: 2026-05-23
|
||||
type: concept
|
||||
tags: [agent, optimization, simplification, meta-learning]
|
||||
sources: [raw/papers/agent-harness-engineering-survey-2026.md]
|
||||
confidence: medium
|
||||
---
|
||||
|
||||
# Adaptive Harness Simplification
|
||||
|
||||
> Harness 设计不应假设**单调地增加更多脚手架**。每个包装器、重置策略、验证器、规划器、记忆规则和权限门都编码了对"模型自身无法可靠完成什么"的假设。随着模型能力变化,Harness 干预应被重新评估而非假定持续有益。
|
||||
|
||||
## 核心洞察
|
||||
|
||||
- Anthropic(2026c):对某个模型有用的上下文重置在新模型上变得可省略,移除它们降低了成本而不降低质量
|
||||
- Bölük(2026b):因子化 model-by-harness 评估可揭示干预何时改善所有模型、仅帮助特定模型家族、或逆转模型排名
|
||||
|
||||
## 元工程议程
|
||||
|
||||
- **Meta-Harness**(Lee et al., 2026):prompt、工具和控制回路可作为优化目标的一部分来搜索
|
||||
- **Natural-Language Agent Harnesses**(Pan et al., 2026):使 harness 模块显式且可消融
|
||||
- 生产系统应向**自适应简化**演进:持续追问哪些控制仍然必要
|
||||
|
||||
## 风险:Benchmark 过拟合
|
||||
|
||||
仅针对狭窄套件自我优化的 Harness 可能变得脆弱。更持久的目标是**自适应简化**:随着任务、工具和模型能力变化持续追问控制必要性。
|
||||
|
||||
## 相关概念
|
||||
|
||||
- [[cost-quality-speed-trilemma]] — 简化是降低成本的一条路径
|
||||
- [[binding-constraint-thesis]] — 约束瓶颈随模型能力变化
|
||||
- [[agent-harness-engineering-survey]]
|
||||
Reference in New Issue
Block a user