--- title: "Harness 与评估 - 工作流专栏 | 思行致远" canonical: "https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/" dateModified: "2026-09-14T19:22:39+08:00" --- 打开 AI·外脑目录 AI·外脑 [返回专栏首页](https://siyuanailab.com/knowledge/) × 站内搜索 ⌕ 主要栏目 - [AI 核心概念](https://siyuanailab.com/knowledge/ai-core/) - [工作流](https://siyuanailab.com/knowledge/workflows/) - [工具配置与选型](https://siyuanailab.com/knowledge/toolkits/) - [信源与标准](https://siyuanailab.com/knowledge/sources/) - [场景案例 待开发](https://siyuanailab.com/knowledge/scenarios/) - [职业与能力](https://siyuanailab.com/knowledge/career/) 专题 - [从这里开始](https://siyuanailab.com/knowledge/start/) - [获取与安装](https://siyuanailab.com/knowledge/install/) - [建设日志](https://siyuanailab.com/knowledge/ops/) [工作流子目录](https://siyuanailab.com/knowledge/workflows/) 基础设施层 - [浏览器自动化 **5**](https://siyuanailab.com/knowledge/workflows/infrastructure/browser-automation/) - [MCP 工具 **5**](https://siyuanailab.com/knowledge/workflows/infrastructure/mcp-tools/) - [文件系统与本地执行 **5**](https://siyuanailab.com/knowledge/workflows/infrastructure/filesystem-local/) - [API 网关与工具路由 **5**](https://siyuanailab.com/knowledge/workflows/infrastructure/api-tool-routing/) Skill 生产线 - [目标明确 **5**](https://siyuanailab.com/knowledge/workflows/skill-forge/goal-definition/) - [Skill 设计 **5**](https://siyuanailab.com/knowledge/workflows/skill-forge/skill-design/) - [代码实现 **5**](https://siyuanailab.com/knowledge/workflows/skill-forge/implementation/) - [测试评估 **5**](https://siyuanailab.com/knowledge/workflows/skill-forge/testing-evaluation/) Skill 参考库 - [写作与内容 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/writing-content/) - [研究与知识 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/research-knowledge/) - [沟通与协作 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/communication-collaboration/) - [产品与项目 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/product-project/) - [数据与分析 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/data-analysis/) - [创意与设计 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/creative-design/) - [商业与营销 **5**](https://siyuanailab.com/knowledge/workflows/skill-library/business-marketing/) Agent Building 只收录有多步工作流、工作标准、Loop、Harness、工具链或通信机制的真 Agent；不收纯 persona、profile、curated list 或只有 system prompt 的仓库。 - [多 Agent 框架 **5**](https://siyuanailab.com/knowledge/workflows/agent-building/multi-agent-frameworks/) - [工作流编排 **5**](https://siyuanailab.com/knowledge/workflows/agent-building/workflow-orchestration/) - [Sub-agent 模式 **5**](https://siyuanailab.com/knowledge/workflows/agent-building/sub-agent-patterns/) - [Harness 与评估 **5**](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/) - [WF-091 Harness 与评估参考：agent lens](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-agent-lens/) - [WF-092 Harness 与评估参考：claude code hooks multi agent observability](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-claude-code-hooks-multi-agent-observability/) - [WF-093 Harness 与评估参考：HALO](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-halo/) - [WF-094 Harness 与评估参考：litefuse](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-litefuse/) - [WF-095 Harness 与评估参考：www project agent observability standard](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-www-project-agent-observability-standard/) Agent Building · Harness & Evaluation # Harness 与评估 eval harness、trace、回放、benchmark、guardrails 和验收机制。 当前公开卡片 5 张 Harness 与评估词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | WF-091 | [Harness 与评估参考：agent lens](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-agent-lens/) | agent-lens | 一个多会话 Agent 轨迹 harness，可运行 Claude Code/Codex、捕获 ATIF 轨迹、追踪文件 diff、重放/重采样并记录 subagent。 | Harnesstraceobservability | | WF-092 | [Harness 与评估参考：claude code hooks multi agent observability](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-claude-code-hooks-multi-agent-observability/) | claude-code-hooks-multi-agent-observability | 一个 Claude Code hooks 可观测系统，把 12 类生命周期事件通过 Python hook 发到 Bun/SQLite/Vue 实时仪表盘。 | Harnesstraceobservability | | WF-093 | [Harness 与评估参考：HALO](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-halo/) | HALO | Hierarchical Agent Loop Optimizer，用 RLM 分析大量执行 trace，找出 harness 级失败模式，再让编码 Agent 修改 harness 并循环验证。 | Harnesstraceobservability | | WF-094 | [Harness 与评估参考：litefuse](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-litefuse/) | litefuse | 一个开源 LLM engineering 平台，覆盖 trace、prompt management、eval、dataset、playground、API 和自托管部署。 | Harnesstraceobservability | | WF-095 | [Harness 与评估参考：www project agent observability standard](https://siyuanailab.com/knowledge/workflows/agent-building/harness-evaluation/agent-building-harness-evaluation-www-project-agent-observability-standard/) | www-project-agent-observability-standard | OWASP 的 Agent Observability Standard，围绕 inspectable、traceable、instrumentable 定义 AOS、AgBOM、OpenTelemetry/OCSF、MCP/A2A hoo | Harnesstraceobservability |