Skip to main content

8. Evaluation & Benchmarks

Evaluating AI agents is fundamentally different from evaluating static LLMs. Agents must be assessed on their ability to reason, plan, use tools, recover from errors, and complete multi-step tasks — not just generate text.


8.1 Why Agent Evaluation Matters​

Key Metrics​

MetricDescriptionHow to Measure
Success RatePercentage of tasks completed correctlyGround truth comparison
Token CostTotal tokens consumed per taskToken counting
LatencyTime to complete a taskWall-clock timing
Tool AccuracyCorrect tool selection and usageAction log analysis
Error RecoveryAbility to recover from failuresInjected failure tests
Planning QualityEfficiency of task decompositionExpert evaluation

8.2 SWE-bench​

The primary benchmark for software engineering agents.

Overview​

SWE-bench evaluates an agent's ability to resolve real GitHub issues by generating patches.

Variants​

VariantIssuesDescription
SWE-bench Full2,294Complete benchmark from 12 Python repos
SWE-bench Lite300Curated subset for faster evaluation
SWE-bench Verified500Human-verified for reliable evaluation

Evaluation Process​

  1. Input: Agent receives a GitHub issue (description + repo state)
  2. Execution: Agent explores codebase, identifies bug, generates patch
  3. Evaluation: Patch applied, test suite run
  4. Result: Pass if all relevant tests pass

Progress Over Time​

PeriodTop Score (Verified)Key Agent
2024 Q1~4%Early SWE-Agent
2024 Q2~12%SWE-Agent + GPT-4
2024 Q3~22%Agentless, AutoCodeRover
2024 Q4~33%OpenAI, Anthropic agents
2025 Q1~42%Claude Code, Codex
2025 Q2~48%Multi-agent approaches
2026 Q1~53%Continued improvement
Official Leaderboard

For the latest scores, see swebench.com


8.3 WebArena & OSWorld​

WebArena​

Evaluates agents on realistic web interaction tasks.

Key Features:

  • 812 web tasks across 5 web applications
  • Tasks include: information retrieval, form filling, navigation
  • Fully reproducible web environment
  • Realistic, open-ended tasks
CategoryExample Tasks
Information Finding"Find the cheapest 4-star hotel in NYC"
Form Filling"Submit an expense report for $50 lunch"
Navigation"Go to Settings → Privacy → Change to Friends Only"
Data Entry"Create a new contact with the following details"

OSWorld​

Evaluates agents on real desktop OS tasks.

Key Features:

  • 369 tasks across Ubuntu, Windows, macOS
  • Real operating system environments
  • Multi-application workflows
  • File, application, and system operations
MetricWebArena (2025)OSWorld (2025)
Top Agent Score~48%~22%
Human Baseline~78%~72%
Gap30%50%

HyperBrowseComp:多语言多模态浏览压力测试(2026-10)​

arXiv 2610.03574 提出的浏览 agent 基准:423 道由母语者手工编写并人工验证的题目,覆盖 13 种语言,题目刻意设计为"极端困难"——需要跨页面、跨模态(文本 + 视觉)的信息综合才能作答。价值在于打破"英语 + 单页检索"的评测惯性:多语言环境下的导航、表单交互与信息抽取能力是浏览 agent 全球可用的前提,而现有 agent 在非英语任务上的衰减长期以来缺少系统性度量。

LiveMACE:演化市场中的过程感知 Agent 评测(2026-10)​

arXiv 2610.09872 指出:在持续演化的环境里,结果(outcome)是 agent 行为与环境变化的闭环产物,仅按结果打分会掩盖真正产生结果的能力。LiveMACE 在演化市场场景中按过程评估 agent 能力——与 SWE-bench 类"最终测试通过"范式形成互补,代表评测社区从"结果导向"向"过程导向"的新趋势。

核心特点:

  • 环境本身随 agent 行为演化,静态基准的"固定题库"假设失效
  • 评估粒度从轨迹终态下沉到中间决策过程
  • 与 HyperBrowseComp 同期出现,共同指向"评测要度量能力而非运气"

来源:arXiv:2610.09872(2026-10-07)

SciExam for ENSO:开放式科研 Agent 的实测式评测(2026-10)​

arXiv 2610.10513 针对开放式科研 agent 评测的根本难题:已知答案、评分表、LLM 评审都无法判断 agent 是否做出了新科学。SciExam for ENSO 借鉴数值天气预报的独立复验思路——让 agent 构建气候模型预测 ENSO 指数,用未来实测数据验证——把"科研能力评测"从问答式推向实测式。

来源:arXiv:2610.10513(2026-10-07)

评测的统计稳定性:同查询不同答案(2026-10-12)​

arXiv 2610.13823(Same Query, Different Answer)测了一个被整个评测范式默认忽略的量:语义查询引擎(用 LLM 逐行判断的查询系统)在重复执行时的 run-to-run 方差。即使温度为 0,重新执行一次评测也可能显著改变结论——基准分数不再是"一个数",而是一个分布。对 leaderboard 文化的直接冲击是:在没有报告方差的情况下,几个百分点的排名差异可能只是噪声。

配套的还有 arXiv 2610.13821(Can Language Models Learn When to Abstain?):让模型学会"什么时候说不知道"——比较训练模型本身 vs 训练独立的答案校验组件两种路线,把弃权决策(abstention)当作可学习、可校准的能力来评测,而不是模型的自然属性。这与决策模型的温度缩放校准同一脉络:置信度必须成为一等公民的评测对象。

来源:arXiv:2610.13823、arXiv:2610.13821(2026-10-09)

长时程评测的执行失败维度与多 Session 工作流(2026-10-13)​

  • MAGMA-Bench(arXiv 2610.13534):现有具身 agent 基准大多假设动作可靠执行,MAGMA-Bench 专门评测动态任务 + 执行失败下的长时程具身 agent——维护任务相关信息、协调异构机器人、随执行展开自适应决策。执行失败(抓空、超时、机械误差)从"噪声"变成一等评测维度。
  • Humanize Olympiad Agents(arXiv 2610.13632):把奥赛级问题求解拆成多 Session 工作流——Solve、Review、Wiki(沉淀为可复用知识库)、Verify(Lean 形式化验证)四个角色协作。这是继 AlphaProof Nexus 之后"LLM 推理 + 形式化验证"路线向通用工作流下沉的又一例:验证环节从一次性登 Science 的系统工程,变成可复制的工作流模板。

两条线的公共指向:评测对象正从"单轮任务成功率"扩展到跨 Session 的知识沉淀质量与失败条件下的鲁棒性——都是真实部署里决定可用性的维度。

来源:arXiv:2610.13534、arXiv:2610.13632(2026-10-09)


8.4 General Agent Benchmarks​

GAIA (General AI Assistants)​

Evaluates general AI assistant capabilities across difficulty levels.

LevelTasksDescription
Level 1153Straightforward, single-step
Level 2251Multi-step reasoning, tool use
Level 396Complex, multi-modal, long-horizon

AgentBench​

Multi-dimensional agent evaluation across 8 environments.

τ-bench​

Evaluates agents on realistic customer service tasks with policy compliance.

  • Tests following complex policies
  • Evaluates tool usage accuracy
  • Measures conversation quality
  • Includes adversarial user scenarios

Claw-Eval-Live(2026)​

一个面向真实工作流的动态 Agent 基准测试,与传统冻结任务集不同,Claw-Eval-Live 的任务会随时间演化。

核心特点:

  • 活基准:任务集持续更新,反映真实工作流的变化
  • 端到端验证:不仅评估最终响应,还验证任务是否真正被执行
  • 多工具覆盖:跨越软件工具、业务服务和本地工作区
  • 可演化性:支持新工作流场景的动态添加

这代表了 Agent 评估的新范式——从静态 benchmark 向动态、可演化的评估体系转变。

来源:arXiv 2604.28139(2026-04-30)

Synthetic Computers at Scale(2026)​

微软研究院提出的大规模合成计算机方法,用于长时程生产力模拟。核心思路是创建模拟的用户计算机环境(含文件夹层级和内容丰富的文档),然后运行跨越 8+ 小时、平均 2000+ 轮次的 Agent 模拟。

核心创新:

  • 1,000 个合成计算机环境:每个都有真实的目录结构和文档(PPT、Excel 等)
  • 双 Agent 架构:一个 Agent 创建任务目标(模拟一个月的人类工作量),另一个 Agent 执行这些任务
  • 规模化潜力:理论上可扩展到数百万甚至数十亿个合成用户世界

实验结果:

  • 每次运行平均超过 8 小时 Agent 运行时间
  • 在域内和域外生产力评估中都产生了显著的性能提升
  • 为 Agent 自我改进和 Agentic RL 提供了基础数据层

这为 Agent 训练开辟了新范式——不再依赖人工标注的轨迹数据,而是通过大规模合成环境自动生成学习信号。

来源:arXiv 2604.28181(2026-04-30)


8.5 LLM-as-a-Judge​

Using LLMs to evaluate agent outputs when human evaluation is impractical.

How It Works​

Evaluation Approaches​

ApproachDescriptionBest For
Single ScoreRate output on 1-5 scaleQuick assessment
Pairwise ComparisonCompare two outputsRelative ranking
Reference-BasedCompare to ground truthTask completion
Multi-CriteriaScore on multiple dimensionsDetailed analysis
Chain-of-Thought JudgeJudge explains reasoningReliability

Best Practices​

  1. Use strong models as judges (GPT-4o, Claude Opus)
  2. Provide clear rubrics with specific criteria
  3. Calibrate with human evaluation on a subset
  4. Use multiple judges for high-stakes evaluation
  5. Randomize presentation order for pairwise comparisons
  6. Track inter-annotator agreement between LLM and human judges

Example Judge Prompt​

You are evaluating an AI agent's response.

Task: {original_task}
Agent Response: {agent_response}

Evaluate on these criteria (1-5 scale each):
1. Task Completion: Did the agent fully complete the task?
2. Accuracy: Is the information correct?
3. Tool Usage: Were tools used appropriately?
4. Efficiency: Was the approach efficient?
5. Clarity: Is the response clear and well-structured?

Provide a score for each criterion and overall assessment.

8.6 Building an Evaluation Pipeline​

Architecture​

Implementation Checklist​

  • Define clear success criteria for each task type
  • Create diverse test set covering edge cases
  • Implement action logging for all agent steps
  • Set up automated evaluation with ground truth
  • Add LLM-as-Judge for qualitative assessment
  • Establish baseline with human performance
  • Track metrics over time (regression testing)
  • Include cost and latency metrics
  • Regular calibration with human evaluators

8.7 Evaluation Frameworks & Tools​

ToolTypeDescription
LangSmithPlatformTracing, evaluation, and testing for LangChain
PromptfooCLIPrompt evaluation and comparison
RagasLibraryRAG-specific evaluation metrics
DeepEvalLibraryLLM evaluation framework
Arize PhoenixPlatformLLM observability and evaluation
BraintrustPlatformEvaluation and experiment tracking

8.8 Evaluation Costs: The Emerging Bottleneck (2026)​

Evaluation is becoming prohibitively expensive — in some cases rivaling training costs:

BenchmarkCostDetails
HAL (Agent Leaderboard)~$40,00021,730 agent rollouts across 9 models
GAIA (single frontier model)$2,829Per run, before caching
MLE-Bench (full sweep)~$100,00075 competitions × 3 seeds × 6 models
HELM (30 models)~$100,000Aggregate API + GPU costs
The Well (scientific ML)960 H100-hoursSingle architecture evaluation

Why Agent Evals Are Especially Expensive​

  • Scaffold sensitivity: Exgentic's $22K sweep found a 33× cost spread on identical tasks just from scaffold choice
  • Non-compressibility: Unlike static benchmarks (MMLU compressible from 14K to 100 items with 2% error), agent benchmarks are noisy and scaffold-sensitive
  • Repeated runs: Adding reliability requires multiple runs, multiplying costs
  • Inference scaling: UK-AISI scaled agentic steps into millions to study inference-time compute

Cost Reduction Strategies​

  1. Flash-HELM approach: Run cheap evaluations first, spend high-resolution compute only on top candidates (100-200× reduction)
  2. Anchor Points: 1-30 examples can rank-order 87 model/prompt pairs on GLUE
  3. tinyBenchmarks: Item Response Theory to select minimal representative subsets
  4. Caching: Reuse rollouts across evaluations where possible

Industry implication: As models improve, evaluation becomes the bottleneck. Teams must budget for eval compute alongside training compute.

🔗 参考:HuggingFace: AI evals are becoming the new compute bottleneck

Open Agent Leaderboard(2026)​

HuggingFace 联合 IBM Research 推出的开放 Agent 评估排行榜,旨在为 AI Agent 的实际能力提供标准化、可复现的评测基准。

核心特点:

  • 开放评估:所有评测任务公开透明,支持社区贡献和复现
  • 多维度覆盖:涵盖工具调用、推理、多轮对话等 Agent 核心能力
  • 企业级参与:由 IBM Research 和 HuggingFace 共同维护,兼顾学术严谨性与工业实用性

这标志着 Agent 评估从学术界走向产业标准化的关键一步。随着 Agent 在生产环境中广泛部署,标准化评估将成为选择和优化 Agent 的基础。

来源:HuggingFace Blog: The Open Agent Leaderboard(2026-05-18)

ITBench-AA:企业 IT Agentic 任务基准(2026-05)​

Artificial Analysis 联合 IBM Software Innovation Lab 发布 ITBench-AA,首个面向企业 IT Agentic 任务的基准测试系列,首期聚焦 Site Reliability Engineering(SRE)任务。

核心发现:

  • 所有前沿模型得分均低于 50%,ITBench-AA SRE 成为最低饱和度的 Agentic 基准之一
  • Claude Opus 4.7(Adaptive Reasoning, Max Effort)以 47% 领先,GPT-5.5(xhigh)46%,Qwen3.7 Max 42%
  • 开源权重模型中,GLM-5.1(Reasoning)以 40% 领先,与 Gemini 3.5 Flash(high)持平
  • Turn 数量差异近 3 倍,更长轨迹并不等于更高准确率——GPT-5.5 平均 31 轮达 46%,Gemini 3.1 Pro Preview 平均 83 轮仅 30%

测试范围:

  • 59 个 SRE 任务(40 公开 + 19 保留),覆盖 Kubernetes 事故响应
  • Agent 需读取日志、追踪依赖关系、在复杂基础设施中识别根因实体
  • 后续将扩展至 Financial Operations(FinOps)和 CISO 任务

这表明尽管前沿模型在编码基准上表现优异,但在真实企业 IT 运维场景中仍有巨大提升空间。

来源:HuggingFace Blog: ITBench-AA(2026-05-27)

OpenSCAD Architectural 3D LLM Benchmark(2026-05)​

ModelRift 发布的 OpenSCAD 建筑建模基准,测试 LLM 生成复杂 3D 建筑模型的能力。参赛模型包括 Codex 5.5 High、Claude Sonnet、Claude Opus、Cursor Composer、Google Antigravity 和 ModelRift。

核心特点:

  • 任务要求 LLM 用 OpenSCAD 代码重建帕特农神庙(Pantheon)等复杂建筑
  • 评估维度包括几何精度、代码可执行性、建筑细节还原度
  • Google Antigravity 2.0 在此次基准中排名第一

这标志着 LLM 编程评估从纯文本/代码生成扩展到 3D 空间理解和参数化建模领域,对 Agent 的空间推理能力提出更高要求。

来源:ModelRift: OpenSCAD LLM Benchmark(2026-05)

HarnessAPI:统一流式 API 与 MCP 工具(2026-05)​

arXiv 论文 2605.22733 提出 HarnessAPI 框架,解决 Python 函数需要同时维护 HTTP 端点(面向人类/CI)和 MCP 工具注册(面向 Agent 运行时)的双重维护问题。

核心理念:Skill-First 设计范式——每个 Python 函数只需定义一次,自动同时暴露为流式 HTTP API 和 MCP 工具。

来源:arXiv:2605.22733(2026-05-21)

DeepWeb-Bench(2026-05)​

由 Sixiong Xie 等人提出的深度研究(Deep Research)基准测试。当前前沿 LLM 的深度研究能力已使既有基准饱和,DeepWeb-Bench 专注于需要大规模跨源证据收集和长链条推导的复杂任务。

核心特点:

  • 要求 Agent 在开放网络上搜索、收集证据并经过扩展推理得出答案
  • 任务设计远超现有基准的难度,专门测试多源信息整合能力
  • 覆盖需要综合多个独立信息源才能回答的深度研究场景

来源:arXiv:2605.21482(2026-05-20)


8.9 Key Takeaways​

  1. Agent evaluation is multi-dimensional — not just text quality
  2. SWE-bench is the standard for coding agent evaluation
  3. WebArena and OSWorld test GUI interaction capabilities
  4. LLM-as-a-Judge enables scalable but approximate evaluation
  5. Always combine automated + human evaluation for reliable results
  6. Track cost and latency alongside quality metrics

Start Simple

Begin with task completion rate as your primary metric. Add more dimensions (cost, latency, tool accuracy) as your evaluation matures.

Benchmark Selection

Choose benchmarks that match your use case:

  • Coding agents → SWE-bench
  • Web agents → WebArena
  • Desktop agents → OSWorld
  • General agents → GAIA / AgentBench