Multiple file types多种文件
The input is a workspace.输入是一整个 workspace。
Files, folders, tables, documents, exports, instructions and local state all matter.文件、目录、表格、文档、导出记录、任务说明和本地状态都可能是线索。
Can an agent turn a messy workspace into a usable, verifiable deliverable? Agent 能不能把一组杂乱的办公文件，整理成可用、可核查的交付物？
Compare the overall score, deterministic Rule checks, and evidence-grounded LLM Judge score.对比总分、确定性的 Rule 检查得分，以及基于证据的 LLM Judge 得分。
Note Model settings: reasoning effort defaults to high, and each model runs with its provider-default inference hyperparameters.
Harness settings: the context window is unified at 200k with a unified auto-compaction threshold; the WebSearch and AskUserQuestion tools are disabled.
Evaluation is pinned to specific harness builds — codebuddy-code:2.109.3 and claude-code:2.1.187; metrics may shift as harness versions evolve.
The HY (Hunyuan) endpoint in this evaluation is served first-party by its provider; the other models are accessed through third-party serving endpoints, whose parameter configuration and request handling may affect metrics.
备注 模型设置：reasoning effort 默认为 high，各模型使用其默认推理超参。
harness 设置：上下文窗口统一为 200k，自动压缩阈值统一；禁用 WebSearch 与 AskUserQuestion 工具。
评测固定在特定 harness 构建上：codebuddy-code:2.109.3 与 claude-code:2.1.187；随 harness 版本演进，指标可能变化。
本次评测中 HY（混元）端点由其提供方第一方部署；其余模型通过第三方推理端点接入，第三方的参数配置与请求处理可能影响指标。
Office evaluates whether an agent can complete everyday workflows inside an offline workspace. The agent reads spreadsheets, PDFs, documents, JSON exports, Markdown notes, file trees and local state, then reconciles, edits, archives, migrates, summarizes or updates them into a usable deliverable.
Reliable execution depends on clear grounding: each claim traces to a source, file changes are accounted for, state and reports agree, and the final artifact is useful to the next person.
Office 考察 Agent 能否在一个离线 workspace 里完成日常办公工作流。Agent 需要读取 Excel、PDF、文档、JSON 导出、Markdown 说明、目录结构和本地状态，再完成对账、编辑、归档、迁移、汇总或状态更新，最终交付可用的产物。
可靠执行需要清楚的依据：每个结论都能追溯到来源，文件改动有据可查，状态和报告保持一致，交付物也能让下一个人直接使用。
Files, folders, tables, documents, exports, instructions and local state all matter.文件、目录、表格、文档、导出记录、任务说明和本地状态都可能是线索。
Completion includes the expected files, state changes and handoff material.完成任务需要留下该生成的文件、该更新的状态和该交接的材料。
The verifier looks at files, state, evidence, structure and reader-facing quality.Verifier 会检查文件、状态、证据链、结构，以及最终内容对读者是否有用。
A typical task is a small work packet. The agent reads the workspace, links evidence across files, makes the requested changes, and leaves behind artifacts that can be checked after the run.一道典型任务就是一个小型工作包。Agent 需要读完 workspace，跨文件串起证据，按要求完成修改或生成，并留下可供检查的产物。
Office distributes 50 tasks across six common work scenes. The scene groups are distribution-informed and describe benchmark coverage — they do not estimate production request traffic. Select a scene to compare its relative reference weight and task count, then explore representative tasks.
Office 的 50 个任务覆盖六类常见办公场景。场景分组为分布对齐的基准覆盖描述，并非对线上请求流量的估计。选择任一场景，即可对照其相对参考权重和任务数量，并查看代表任务。
Note The two distributions use the same scene order. Gray compares each scene's relative coverage share in the distribution-informed reference; green compares Office task counts, normalized to the largest scene in each distribution. Selecting either row highlights both and moves the arrow from the Office row to its representative tasks. 备注 两组分布采用相同的场景顺序。灰色比较各场景在分布对齐参考中的相对覆盖占比，绿色比较各场景的 Office 任务数量；进度条分别按各自分布中的最大值归一化。点击任一行会同步高亮两处分布，并由 Office 任务行的箭头指向右侧代表任务。
Rules verify concrete artifacts; LLM Judge evaluates evidence-grounded quality. Rules cover required files, parseability, schemas, extracted content, workspace changes, source binding and safety boundaries. LLM Judge evaluates semantic qualities such as a useful handoff and reasoning that stays consistent with the evidence.Rule 核对具体文件和结果，LLM Judge 评估有证据支撑的质量。 Rule 检查文件是否存在、能否解析、schema 和抽取内容是否正确、workspace 是否按要求变化、来源是否清楚、安全边界是否守住。LLM Judge 评估交接说明是否有用、理由是否和证据一致。
Scripts check output contracts, schemas, parsed values, file diffs, state changes and evidence links.脚本会检查输出契约、schema、解析值、文件差分、状态变化和证据链接。
Rubrics look at policy fit, exception handling, reasoning, reader usefulness and consistency with the submitted evidence.Rubric 会看任务口径、异常处理、推理依据、读者可用性，以及内容是否和提交证据一致。
An Office score answers one question: did the agent finish the work in a way that can be verified? Grading uses the artifacts left in the offline workspace — files, state changes, evidence links and handoff material. The three views below clarify that scope.
Office 的分数回答一个问题：Agent 有没有把工作做完，并留下可核查的结果？评分依据是离线 workspace 中最终留下的文件、状态变化、证据链接和交接材料。下面三个角度说明这项分数覆盖的范围。
The score reflects completed files, state changes, reports and handoff artifacts.分数反映文件、状态变化、报告和交接产物是否完整交付。
Tasks run in reproducible workspaces that resemble business systems while remaining independent of live services and private accounts.任务在可复现的 workspace 中运行，保留业务系统的工作形态，同时独立于真实服务和私人账号。
Checks cover file contracts, content, state, source links, safety boundaries and handoff quality.核查覆盖文件契约、内容、状态、来源链接、安全边界和交接质量。