AI Practices AI实践 2d ago Updated 2d ago 更新于 2天前 46

Evaluating AI Agent Skill Performance with NVIDIA SkillEvaluator 使用 NVIDIA SkillEvaluator 评估 AI Agent 技能性能

NVIDIA SkillEvaluator is an open-source evaluation framework that measures the impact of verified AI agent skills through a three-tier process: static safety checks, distinctiveness analysis, and live task runs in isolated sandboxes Benchmarking over 300 verified skills across 30+ NVIDIA products showed average Skill Lift gains of 31 points overall and 39 points excluding Security across Correctness, Discoverability, Effectiveness, and Efficiency dimensions Baseline agent performance without ski NVIDIA SkillEvaluator通过三层评估流程(静态检查、独特性分析、隔离环境实时任务运行)量化已验证技能对AI代理性能的影响。 基准测试覆盖300多个已验证技能和30多个NVIDIA产品,显示在正确性、可发现性、有效性和效率方面平均提升31分(排除安全因素后为39分)。 结果表明产品领域和评估设计对Skill Lift的影响大于代理harness选择,且token或执行节省因技能而异,需针对性优化。 SkillEvaluator是开源工具,支持通过Harbor框架在隔离沙箱中进行有/无技能的对比评估,生成Skill Lift分数。 当前代理无技能辅助时基线得分较低(正确性46、可

62
Hot 热度
65
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • NVIDIA SkillEvaluator is an open-source evaluation framework that measures the impact of verified AI agent skills through a three-tier process: static safety checks, distinctiveness analysis, and live task runs in isolated sandboxes
  • Benchmarking over 300 verified skills across 30+ NVIDIA products showed average Skill Lift gains of 31 points overall and 39 points excluding Security across Correctness, Discoverability, Effectiveness, and Efficiency dimensions
  • Baseline agent performance without skills was surprisingly low (39-46 out of 100), indicating significant room for improvement in how agents utilize tool documentation and guidance
  • Product domain and evaluation design influence Skill Lift more than the choice of agent harness, and token/execution savings vary considerably by individual skill
  • Skills are distributed across multiple platforms including Claude Code, Codex, Cursor, Skills.sh, ClawHub, and Hermes Hub

Why It Matters

This is highly relevant to AI practitioners building agent systems, as it provides the first systematic, open evaluation framework for measuring whether skills actually improve agent performance rather than assuming they do. The finding that baseline agent performance without skills is only 39-46 out of 100 demonstrates that even capable models struggle significantly with tool discovery and correct usage, making skill evaluation critical for production deployments. The methodology also offers a replicable blueprint for organizations looking to validate their own agent skill implementations.

Technical Details

  • Three-tier evaluation architecture: Tier 1 performs static checks including schema validation, quality scoring, security scanning for prompt injection and data exfiltration, PII detection, license checks, and script linting. Tier 2 uses embedding similarity to detect duplicated guidance within skills and overlapping coverage across the catalog. Tier 3 runs live evaluations in isolated Docker sandboxes using Harbor framework, comparing agent performance with and without each skill installed.
  • Skill Lift measurement: Calculated as the difference between with-skill and without-skill scores across five dimensions—Correctness (final answer accuracy), Discoverability (right skill loading relevance), Effectiveness (goal completion and workflow adherence), Efficiency (minimal wasted steps and redundant tool calls), and Security (no regression from skill installation).
  • Benchmark scope: Over 300 verified skills evaluated across 30+ NVIDIA products using two independent agent harnesses (Codex and Claude Code), with macro-averaged scores giving each skill-harness pair equal weight.
  • Open-source tooling: SkillEvaluator integrates with Harbor for isolated execution, supports --env-mode docker for sandboxing, and generates evaluation datasets with explicit, implicit, contextual, and negative test cases via skillevaluator create-eval-dataset.
  • Distribution ecosystem: Verified skills are published as packaged, signed capability descriptors and distributed through multiple channels including Claude Code, Codex, Cursor plugins, Skills.sh, ClawHub, and Hermes Hub.

Industry Insight

  • The low baseline scores (39-46) suggest that most current agent deployments significantly underperform without structured skill guidance, making skill evaluation and curation a high-priority investment for any organization building production agent systems
  • The finding that domain and evaluation design matter more than harness choice implies that skill quality is portable across agent platforms, reducing vendor lock-in concerns and enabling a more interoperable skills ecosystem
  • Organizations should adopt continuous evaluation practices rather than one-time validation, as the NVIDIA catalog is evaluated continuously and Skill Lift varies by skill—targeted optimization of individual skills will yield better ROI than blanket skill deployment

TL;DR

  • NVIDIA SkillEvaluator通过三层评估流程(静态检查、独特性分析、隔离环境实时任务运行)量化已验证技能对AI代理性能的影响。
  • 基准测试覆盖300多个已验证技能和30多个NVIDIA产品,显示在正确性、可发现性、有效性和效率方面平均提升31分(排除安全因素后为39分)。
  • 结果表明产品领域和评估设计对Skill Lift的影响大于代理harness选择,且token或执行节省因技能而异,需针对性优化。
  • SkillEvaluator是开源工具,支持通过Harbor框架在隔离沙箱中进行有/无技能的对比评估,生成Skill Lift分数。
  • 当前代理无技能辅助时基线得分较低(正确性46、可发现性42、有效性39、效率39),凸显技能工程对提升代理性能的关键作用。

为什么值得看

这篇文章为AI从业者提供了系统化的技能评估方法,帮助量化技能对代理性能的实际影响,从而优化代理工作流程。基准测试结果揭示了当前代理在没有技能辅助下的性能瓶颈,为技能开发和验证指明了改进方向。

技术解析

  • 三层评估流程:Tier 1进行静态检查(模式验证、安全扫描、PII检测等);Tier 2通过嵌入相似度分析技能内部重复指导和跨目录覆盖重叠;Tier 3在隔离沙箱中运行实时任务,比较有/无技能时的代理表现。
  • Skill Lift计算:基于Harbor框架,每个评估用例在相同提示、模型和任务输入下运行两次(有技能和无技能),得分差值即为Skill Lift(以点为单位),确保控制变量。
  • 基准测试规模:评估300多个已验证技能,覆盖30多个NVIDIA产品,使用Codex和Claude Code两个独立代理harness,得分按技能-harness对宏观平均。
  • 评估数据集:通过skillevaluator create-eval-dataset生成,包含显式、隐式、上下文和负面案例,每个案例有ID、提示、预期输出和可选断言。
  • 基线性能:无技能时平均得分较低(正确性46、可发现性42、有效性39、效率39),安全基线为97,主要验证技能安装不引入安全回归。

行业启示

  • 技能工程成为关键领域:代理性能高度依赖上下文质量,标准化技能评估流程将推动技能开发从经验驱动转向数据驱动,提升代理可靠性和效率。
  • 评估设计比代理选择更重要:产品领域和评估设计对Skill Lift的影响大于代理harness,建议企业优先优化技能内容和评估方法,而非盲目更换代理框架。
  • 针对性优化可提升资源效率:token或执行节省因技能而异,需根据具体技能进行精细化调整,以降低运营成本并提高代理响应速度。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Evaluation 评测 GPU GPU Product Launch 产品发布 Benchmark 基准测试