Office agent research in China
Historical report. Information through September 9, 2026; some dynamic documents were checked on September 10. Public-source research without installed-product testing. These versions are not a controlled with/without-toolkit experiment.
Full report (English translation) · Interactive case (Chinese)
Direct prose
The opening of the workspace section moves from a broad value claim to how the products organize work.
Replace rhetorical setup with actions and comparison criteria.
The final report shortens the rhetorical setup and a vague closing, then names tools, context, and device availability.
Changes
Initial draft · Translated excerpt
The value of independent office agents lies not in moving a chat box onto the desktop, but in bringing materials, execution, and revisions into one stretch of work. For people who frequently organize files from different sources, this approach can reduce uploading, downloading, and switching tools. Yet keeping files on a computer does not mean tasks run continuously or all processing happens offline. The differences between Kimi Work and Coze show that independent workspaces do not share a single implementation.
The opening frames value through a rhetorical contrast; the closing only says implementations differ.
Final report · Translated excerpt
Independent office agents organize materials, execution, and revisions in one workspace. For people who frequently organize files from different sources, this approach can reduce uploading, downloading, and switching tools. Continuous operation and the location of data processing still depend on the implementation. The differences between Kimi Work and Coze center on who supplies tools, who retains context, and who keeps the device online.
The report states the mechanism directly and identifies comparison criteria, while retaining its conditions.
What changed
- Shortened：Remove the rhetorical setup and state how materials, execution, and revisions share a workspace.
- Made specific：Turn “implementations differ” into tools, context, and device availability. Keep the conditional benefit of less file handling.
Reasoning
Which wording adds little
Moving a chat box is not the mechanism examined in this paragraph. Using it as a contrast delays the point. Saying implementations differ gives the reader no criteria.
Why this revision helps
The final paragraph states the workspace mechanism and identifies three conditions. Readers can compare the products on tools, context, and device availability.
Decision and tradeoffs
- Chosen：Express the judgment through actions, objects, and conditions, with less rhetorical setup.
- Retained：Less file handling is a possible benefit; continuous operation and processing location still depend on the implementation.
- Not inferred：This edit neither bans contrastive sentences nor identifies whether a passage was written by AI.
Initial passage (Chinese) · Final passage (Chinese) · Author editing record (Chinese)
Review
- Initial and later drafts：The paragraph opened with a value contrast and ended by saying implementations differ. Early retained drafts kept that wording.
- Author editing check：The author found that repeated limitations and verification language displaced substantive product analysis.
- Rewritten intermediate draft：The next retained draft stated the workspace mechanism and named tools, context, and device availability.
- Final report：The final report retained that revision, including the conditional benefit and operating constraints.
The passage change is verifiable in the manuscripts. Its editing context is an author check, not a model’s sentence-by-sentence instruction.
Intermediate passages and author check (Chinese)
Process narration
The opening drops an outline announcement and brings essential source and scope information together.
Remove the outline announcement; keep the research limits.
The chapter headings already provide navigation. Public-source research and the absence of installed-product testing still matter to how readers interpret the findings.
Changes
Initial draft · Translated excerpt
The following sections cover work objects, independent workspaces, office suites, desktop execution, and commercialization. The evidence consists of public product documents and examples, without installed-product testing; a single use record does not represent industry performance.
After the table, the draft announced its structure and added source and testing limits.
Final report · Translated excerpt
The research uses public documents and examples, without installed-product testing. Some dynamic descriptions were checked on September 10; their capabilities are not attributed retrospectively to earlier versions.
The announcement was deleted. Source and testing limits moved to the opening; this is a reorganized excerpt, not a replacement of the same sentence.
What changed
- Removed：Drop the section preview and the earlier announcement that three judgments should be stated first.
- Moved and retained：Keep source, testing, and sample limits in the opening. The final report also identifies the dates of dynamic-source checks.
Reasoning
What counts as a writer’s aside
The section preview describes the author’s plan without adding product information. Headings already show the structure. Announcing that conclusions are about to follow also delays those conclusions.
Why the limits remain
Public documents, no installed-product testing, and the limits of a single example affect how readers interpret the findings. They remain in the report after the outline narration is removed.
Decision and tradeoffs
- Chosen：Let the product findings carry the opening, without a repeated outline announcement.
- Retained：Scope, historical dates, sources, testing limits, and sample representativeness.
- Display limit：This example deletes an outline announcement and relocates essential scope information. The excerpts come from different positions in the same report.
Initial outline passage (Chinese) · Final scope passage (Chinese) · Author editing record (Chinese)
Review
- Initial draft：The opening announced three judgments and later previewed the section order. Scope information appeared in several places.
- Author editing check：The author called for leading with supported product changes and judgments, while reducing repetition and retaining material uncertainty.
- Reorganized intermediate draft：The next retained draft removed both announcements and grouped source and scope information after the main findings.
- Final report：The final report kept that arrangement and added dynamic-source dates and sample limits.
The structure change is verifiable in the manuscripts. Its editing direction is an author check, not an independent model finding.
Intermediate passages and author check (Chinese)
aily and Doubao Work
Feishu introduced aily custom agents in April; its late-August article used the name Doubao Work. This comparison concerns the relationship between those products.
The retained sources do not establish that Doubao Work inherits from aily.
Two product descriptions can be analyzed on their own. Their dates and similar direction do not establish a version lineage.
Changes
Initial draft · Translated excerpt
This is evidence of continuity between the current product description and the earlier feature direction
The same paragraph already cautioned against assuming a one-to-one transfer of features and names.
Final report · Translated excerpt
The two points in time show product directions in object operations and organizational-context collaboration; they do not establish version inheritance between aily and Doubao Work
The migration caveat remains; the report now explicitly limits what the two sources establish.
What changed
- Narrowed：“Evidence of continuity” becomes product directions at two points in time, without a claim of version inheritance.
- Retained：Both official product descriptions and the warning against carrying earlier capabilities over to the later product.
Reasoning
What each source supports
The April source describes aily operating on documents, multidimensional tables, and tasks. The late-August article describes Doubao Work using organizational context and returning results to collaboration.
Why the inference was narrowed
Similar functions and successive dates do not establish renaming, version inheritance, or feature migration. The author’s follow-up did not obtain that evidence, so wording implying continuity was removed.
Decision and tradeoffs
- Chosen：Analyze each source without joining them into a verified product lineage.
- Tradeoff：The product relationship stays unresolved. Comparisons must use capabilities supported by each source separately, without merging their feature lists.
- What could change it：An explicit renaming notice, version mapping, or feature-migration document would support revisiting the relationship.
Feishu aily source · Doubao Work source
Review
- Initial and later drafts：“Evidence of continuity” remained alongside the migration caveat. The wording implied a stronger relationship than the available sources established.
- Review finding：A model review flagged that “evidence of continuity” could imply version inheritance. The author accepted a precision correction.
- Source check and decision：The follow-up obtained no renaming or migration proof. An aily domain in site navigation was not treated as proof either. Both dated product descriptions were kept.
- Intermediate and final report：The next retained draft switched to product directions at two points in time and explicitly rejected an unsupported inheritance claim. The final report kept that wording.
Model-review findings are relayed through the author’s retained decisions.
Intermediate excerpts (Chinese)
Coze’s models
Coze China groups native agents, third-party cloud agents, and local connections in one interface. An execution framework organizes actions; the underlying model handles understanding and reasoning.
Coze supplies models for its cloud-hosted Claude Code.
The report identifies Claude Code and Codex CLI as execution frameworks. A program’s brand does not, by itself, establish which model powers a particular deployment.
Changes
Initial draft · Translated excerpt
Coze brings native agents, third-party agents hosted in the cloud, and local connections into one collaboration interface. The native version includes platform scheduling, email, and cloud resources; the other modes do not automatically inherit those built-in services.
The section distinguished execution modes and services, but did not identify their specific model supply.
Final report · Translated excerpt
Native agents can select models such as Doubao, Kimi, GLM, and MiniMax. Third-party cloud mode runs execution frameworks such as Claude Code and Codex CLI on Coze cloud computers, using models supplied by Coze rather than being tied to the original provider’s account and model.
This paragraph was added to the same product section. Its existing mode and permission analysis remains.
What changed
- Added：Specific native-model choices and the distinction between third-party cloud frameworks and their model supply.
- Retained：The original service boundaries and local-device conditions; the additional paragraph does not make all three modes equivalent.
Reasoning
The fact layer missing from the draft
General discussion of model importance and execution modes cannot answer which models a particular product uses. A reader may also mistake the Claude Code program brand for a model provider.
What the additional sources establish
The product introduction lists native-model choices. The cloud-agent document explains frameworks running on Coze computers with Coze-supplied models. They allow model supply and execution mode to be described separately.
Decision and tradeoffs
- Chosen：Add verifiable product configuration while retaining the distinct roles of models, tools, context, and permissions.
- Not inferred：Configuration documents do not measure a model’s share of time saved or real-task completion rates.
- What is still needed：Comparing model effects requires fixed tasks, inputs, tools, permissions, and result criteria, followed by errors, failed calls, and rework.
Agent types · Product introduction · Third-party cloud documentation
Review
- Initial draft：The Coze section described execution modes and service boundaries. Its conclusion covered general model roles but lacked specific supply details.
- Review finding：Two model reviews found the analysis too general. The author combined their substantive findings and added product facts rather than duplicate paragraphs.
- Expanded intermediate draft：The author returned to official type, product, and cloud documents. The next draft added native choices and cloud-model supply, with product-level synthesis in the conclusion.
- Later decision and final report：A later adjudication added no duplicate mechanism paragraph. Conditions for controlled comparisons were retained; simply switching a model was not treated as a causal experiment.
Model-review findings are relayed through the author’s retained decisions.
Intermediate excerpts (Chinese)
Research request
Use Research Toolkit for this task:
https://github.com/rrrrrredy/research-toolkit
Research office agents available to users in China in 2026.
Compare Kimi Work, Coze, Feishu / Doubao Work, Wukong,
WPS, and WorkBuddy. Write for AI product and workplace leads.
Explain which files and business objects each product changes,
where tasks run, and how models, tools, context, permissions,
and human involvement affect delivery and rework.
Use information available as of [DATE]. Cite important claims.
Distinguish company statements, media experiences, and measured
outcomes. Explain what could overturn your conclusions.
Read SKILL.md. Clarify missing requirements and agree on an
outline before collecting sources.