实
OpenHands
All Hands AI’s open-source multi-agent coding platform — one of the evaluation frameworks in SWE-EVO
OpenHands’ CodeActAgent architecture surfaced a key finding in SWE-EVO: GLM-5 scores 37.5% on SWE-agent but only 8.33% on OpenHands — the same model, yet framework differences drive massive performance divergence. This proves agent capability is a function of model × framework, not an intrinsic property of the model.
Framework Comparison
Framework
Architecture
Characteristics
OpenHands
CodeActAgent
Multi-agent platform, unified action space
SWE-agent
Single agent
Emphasizes agent-computer interface design
Codex
Implicit loop
Cloud sandbox, bidirectional JSON-RPC
LangGraph
Explicit graph
StateGraph nodes and edges
Key Insights GLM-5 37.5% vs. 8.33% The same model differs by 4.5× across frameworks — framework prompt style and interaction patterns decisively shape performance Capability = Model × Framework There is no “intrinsic agent ability” of a model — evaluation results are always framework-specific Up to 100 Iterations The iteration cap in SWE-EVO settings — a resource boundary for long-horizon tasks
→ Implicit Loop Architecture · SWE-Bench · Codex · LangGraphSWE-EVO arXiv:2512.18470
OpenHands
简介
OpenHands 是一个开源多 agent 编码平台，使用 CodeActAgent 架构，支持在多个 benchmark 上评估 AI 编码 agent。由 All Hands AI 团队开发和维护。
在 SWE-EVO 中的角色
OpenHands 是 SWE-EVO 评估中使用的两个 agent 框架之一（另一个是 SWE-agent），配置为 CodeActAgent，最多 100 次迭代。
一个值得注意的发现：某些模型在不同框架上表现差异极大。GLM-5 在 SWE-agent 上 37.5%，在 OpenHands 上仅 8.33%。这说明 agent 能力是模型 x 框架的函数——框架的 prompt 风格和交互模式会显著影响模型表现。
与其他框架的对比
| 框架 | 架构 | 特点 |
|---|---|---|
| OpenHands | CodeActAgent | 多 agent 平台，统一行动空间 |
| SWE-agent | 单 agent | 强调 agent-computer interface 设计 |
| Codex | 隐式循环 | 云端沙箱，双向 JSON-RPC |
| LangGraph | 显式图编排 | StateGraph 定义节点和边 |
相关概念
- Agentic systems — OpenHands 所属的系统类型
- Implicit loop architecture — CodeActAgent 的架构范式
- Harness engineering — 框架即 harness
References
sources/arxiv_papers/2512.18470-swe-evo.md