Coze (ByteDance)

1. Introduction

1.1. Three-Tier Product Structure

Within the ByteDance ecosystem, "Coze (扣子)" is not a single product but a three-tier structure with the same name, which must be distinguished clearly before vendor selection:

The relationship among the three can be summarized as: Coze proves out production first, then returns the engine to developers; Volcano Engine packages the same set of capabilities into an enterprise AI platform.

1.2. Basic Information Card

1.3. Development Timeline

1.4. Position in the AI Harness System

Coze is a typical form of an Agent Platform, and among them it is the most aggressive on "channel distribution + ecosystem openness":

• On top of L1~L3 it adds UI, tenant, and billing, and additionally adds a channel distribution layer (Feishu / WeChat / Douyin / Web / Discord);
• It belongs to the same Agent Platform category as Dify, but with a different strategic focus: Dify emphasizes "a privatizable global open-source platform", while Coze emphasizes "no-code + ByteDance ecosystem + multi-channel publishing + rich built-in plugins";
• By open-sourcing Coze Studio, it extends itself downward from a "platform" into a "privatizably deployable Harness foundation".

2. Glossary

3. Feature Description

3.1. No-Code Agent Building

Coze's core design goal is to let non-technical users such as product managers, operations staff, customer service, and copywriters also build Agents:

• Visual editor: drag nodes, connect steps, configure parameters;
• Multiple built-in LLMs (Doubao, Tongyi Qianwen, GLM, etc.);
• 100+ built-in plugins (web search, send email, read databases, generate images, etc.);
• Official claim: an AI Agent can be built in 5 minutes.

This is highly similar to Dify's low-code route; the difference is that Coze's channel distribution and ByteDance ecosystem integration run deeper.

3.2. Plugin and Skill Ecosystem

Be especially careful about the discrepancies in ecosystem-scale figures: different sources give widely differing numbers such as 100+ and 700+; the source and the point-in-time of each statistic must be noted when citing.

3.3. Persona / Equipment / Skill Three-Layer Model

The core idea proposed in Coze 2.5 is "giving AI a persona":

The "Equipment" item deserves special emphasis: the cloud PC and cloud phone essentially hostify L2's execution environment — Agents no longer call APIs only within the text domain, but own a complete operating-system execution surface. This significantly raises L2's capability ceiling while also amplifying L6's governance difficulty.

3.4. Workflow and Multi-Agent Collaboration

• Workflow: visual drag-and-drop that solidifies Agent logic into reusable steps;
• Agent World (2.5): the "parallel network" of AI Agents, letting different Agents collaborate with each other; supports multi-Agent collaboration and dynamic routing;
• Project Space (3.0): unifies goals, members, Agents, files, and process artifacts into an independent task management space;
• Multi-person multi-Agent collaboration (3.0): supports flexible combinations of "one person + multiple Agents" or "multiple people + multiple Agents";
• Third-party Agent integration (3.0): can integrate local Agents such as Claude Code, Codex CLI, and OpenClaw.

The last item is an important signal of openness: Coze no longer requires Agents to be built within its own ecosystem, but accepts external Agents for integration and collaboration.

3.5. Multi-Channel Publishing

Coze's most distinctive product capability is one-click publishing to multiple channels: Feishu, WeChat, Douyin, Web, Discord, etc.

From a Harness perspective, this is not a "marketing feature", but a structural capability of the L2 Tools and Execution layer — it turns an Agent's "reach channels" into configurable publishing targets, something the other six platforms do not have. For domestic businesses, this is often the decisive selection factor.

3.6. Open-Source Coze Studio and Coze Loop

Coze Studio:

• Apache 2.0, repo coze-dev/coze-studio;
• Extracts all of visual Agent orchestration, RAG, Plugin, Workflow, model routing, and Chat SDK;
• Microservices architecture + domain-driven design, backend Go, frontend React + TypeScript;
• Services include model management, Agent building, workflow execution, and resource management;
• Supports deployment with Docker and Docker Compose, minimum 2-core 4 GB;
• Provides OpenAPI and Chat SDK for external integration.

Coze Loop: the agent evaluation and operations component open-sourced in the same batch, focusing on "Agent iteration" rather than "Agent development".

The two projects — one for development and one for iteration — form a complete closed loop. In the Harness six-layer model this corresponds to L5 (Evaluation and Observability layer), and it is a structural advantage of Coze over Dify (Dify's L5 relies on annotation and third-party integration).

Apache 2.0 means: commercially usable, no license required, customizable as needed, and privately deployable. This contrasts sharply with Dify's non-OSI license — for teams planning multi-tenant commercial use based on the source code, Coze Studio's license path is more permissive than Dify's.

3.7. Enterprise Side: HiAgent / ArkClaw / AgentSphere

Volcano Engine's approach is a "agile + steady" two-punch strategy: ArkClaw covers exploratory personal scenarios, HiAgent covers scaled enterprise scenarios, and AgentSphere provides a unified governance framework.

4. Platform Architecture

图 4-1｜Coze 云侧分层架构：从渠道分发到火山引擎底座

数据来源：基于本文分析绘制的示意图。

4.1 Coze Studio Tech Stack

┌──────────────────────────────────────────────────────────┐
│ 前端：React + TypeScript                                   │
│  可视化拖拽编排 · 调试 · 发布                               │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 后端微服务（Go，领域驱动设计）                               │
│  模型管理服务 · 智能体构建服务 · 工作流执行服务 · 资源管理服务 │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 能力层                                                     │
│  RAG 知识库 · Plugin 系统 · Workflow 引擎 · 模型路由         │
│  OpenAPI · Chat SDK                                       │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 部署：Docker / Docker Compose（最低 2 核 4 GB）              │
└──────────────────────────────────────────────────────────┘

Coze Loop (independent component): evaluation and operations, covering the Agent's iteration loop.

4.2 Cloud-Side Layered Architecture

┌──────────────────────────────────────────────────────────┐
│ 渠道分发层（扣子独有）                                       │
│  飞书 · 微信 · 抖音 · 网页 · Discord                        │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 平台层：扣子（coze.cn / coze.com）                          │
│  零代码搭建 · 项目空间 · 行业专家技能 · 桌面端 / 手机 App      │
│  Agent World（多智能体协作网络）                             │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 能力层                                                     │
│  人格（长期记忆）· 装备（云电脑 / 云手机）· 技能（Skills）      │
│  插件（100+ / 700+）· RAG 知识库 · Workflow                 │
│  第三方 Agent 接入（Claude Code / Codex CLI / OpenClaw）     │
└──────────────────────────────────────────────────────────┘
                          │
┌──────────────────────────────────────────────────────────┐
│ 底座：火山引擎                                              │
│  豆包大模型 · MaaS · ArkClaw / HiAgent / AgentSphere        │
│  累计 tokens 调用 430 万亿 + · 累计 100 万+ 企业客户           │
└──────────────────────────────────────────────────────────┘

4.3 Processing Flow of a Multi-Channel Agent Request

1. A user initiates a request from any channel — Feishu / WeChat / Douyin / Web / Discord;
2. The channel adaptation layer normalizes it into a platform-internal message;
3. Load persona memory (long-term memory system);
4. If a knowledge base is configured: retrieve → inject into context;
5. If a Workflow is configured: execute per the canvas; if plugins / Skills are configured: call the corresponding capability;
6. If equipment is configured: execute operations in the cloud PC / cloud phone;
7. For multi-Agent scenarios: route to other Agents via Agent World (including third-party integrated local Agents);
8. Generate a response and return it to the original channel;
9. Coze Loop collects evaluation and operations data for iteration.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Coze's three main tools in L1:

1. RAG knowledge base: an ingestion-and-retrieval pipeline similar to Dify's, serving as the Agent's factual foundation;
2. Long-term memory system (persona layer): preserves the Agent's identity and memory across sessions;
3. Skills: 365 industry skills loaded on demand — conceptually similar to the progressive-disclosure Skills in the Claude Agent SDK, but whether a progressive-disclosure mechanism is adopted has not been confirmed by public material.

Shortcomings: there is no public context-compression primitive; retrieval-result ordering and trimming strategies are largely unconfigurable; and how the large context (screenshots, DOM) introduced by the cloud PC / cloud phone should be compressed lacks public material.

5.3. L2 Tools and Execution Layer

This is the layer where Coze shows the greatest structural difference among the other six platforms, for three reasons:

1. Hosted execution environment: the cloud PC (Ubuntu) and cloud phone (Android 13) give Agents a complete operating-system execution surface, with a capability ceiling far above "calling APIs";
2. Channels as tools: one-click publishing to Feishu / WeChat / Douyin / Discord turns "reach" into a configurable capability;
3. Dense plugin ecosystem: 100+ built-in, with third parties claiming 700+ in the ecosystem.

The corresponding cost is a sharp rise in L6 difficulty: an Agent that can operate a cloud PC, speak in WeChat, and read/write an enterprise database has a blast radius per mistake far larger than a pure-text Agent.

5.4. L3 Orchestration and Control Layer

Coze's L3 follows a low-code + dynamic collaboration route:

• Workflow provides deterministic orchestration capability;
• Agent World provides multi-Agent dynamic collaboration and routing;
• Project Space unifies goals, members, Agents, files, and process artifacts in an organized way;
• Starting from 3.0, both "one person + multiple Agents" and "multiple people + multiple Agents" collaboration modes are supported.

The tension between flexibility ↔ predictability manifests here as: the more flexible multi-Agent dynamic routing is, the more unpredictable the execution path becomes. Coze 3.0's introduction of Project Space and "process artifacts" integration can be seen as compensation for this tension — trading process visibility for a measure of predictability.

Comparison with Dify: Dify officially recommends "demoting the autonomous loop into a single node within a workflow using the Agent Node" to gain determinism; Coze does the opposite, making the multi-Agent collaboration network (Agent World) a selling point. The two choose different positions on the same axis.

5.5. L4 Memory and State Layer

What exists:

• Persona-layer long-term memory system (cross-session);
• Project Space: integrates goals, members, Agents, files, and process artifacts — a lightweight task-oriented state management;
• Multi-channel session context.

Missing (not seen in public material):

• Run-level checkpoints and crash resume;
• Artifact version management;
• Explicit separation of semantic memory and episodic memory.

Therefore it is rated "Medium-strong" — persona memory and Project Space are distinctive designs, but the depth of engineered state management is below LangGraph's Checkpointer or ADK's three services.

5.6. L5 Evaluation and Observability Layer

Coze Loop is this platform's most important structural asset on L5. Open-sourced in the same batch as Coze Studio and equally under Apache 2.0, it is positioned as the "agent evaluation and operations component", responsible for Agent iteration rather than development.

The strategic implications of this design:

• It turns evaluation from a "platform-internal feature" into an "independently deployable open-source component";
• Even teams that do not use the Coze platform can use Coze Loop alone for Agent evaluation;
• Compared with Dify's "annotation feedback + third-party integration" approach, Coze's evaluation is a first-class citizen component.

But note: no official detailed documentation was found on Coze Loop's specific evaluation capabilities (metric types, regression-set management, trajectory-scoring methods), so it is rated "Medium-strong" rather than "Strong".

5.7. L6 Governance and Security Layer

Assessment: Coze's governance narrative on the enterprise side is complete (agile + steady + digital-employee management paradigm), but there is very little public information on governance mechanisms on the platform side and the open-source side. For vendor selection, this means:

• If governance capability is needed, go with Volcano Engine's enterprise offering rather than self-hosting Coze Studio;
• When self-hosting Coze Studio, L6 largely needs to be built from scratch.

5.8. Concrete Manifestations of Three Inherent Tensions

6. Practical Cases

Case 1: ARR and platform scale (project-verified hard data)

• Coze reached the ARR $100M milestone in early 2026;
• 5–10 million messages consumed daily;
• 10,000+ Machine Learning related registered enterprises;
• Volcano Engine foundation: MaaS + Agent platform cumulative 430+ trillion tokens called, public cloud presence in 50+ industries, cumulative 1 million+ enterprise customers.

Among the seven platforms in this group, this is the only one with all three sets of hard data simultaneously: ARR, daily active message volume, and enterprise customer count.

Case 2: Volcano Engine industry deployment (IDC report and Volcano Engine public claims)

Case 3: HiAgent enterprise Agents (public media reports)

• With Aima Electric Vehicles, built four AI scenarios: 数管家, 晓师傅, 文博士, and 效枢官;
• At Guanghua School of Management, Peking University, the Agent "豆角" assists teachers with lesson preparation and teacher-student interaction;
• Hailiang Group released 150 Agents covering business management, safe production, and smart teaching.

The above industry and customer cases all come from IDC report retellings and Volcano Engine public claims; no third-party independently audited quantitative effectiveness data was found (such as cost-reduction percentages, efficiency improvement multiples, or labor-substitution rates), and this is noted honestly without fabrication.

Parts where no public quantitative data was found: effectiveness data for Coze platform's own customer cases, details of Coze Loop's evaluation capabilities, and enterprise-deployment-scale data for the open-source version — none were found as of the retrieval date 2026-09-12, noted honestly.

7. Summary

7.1. Strengths

1. Unrivaled channel distribution: one-click publishing to Feishu / WeChat / Douyin / Discord is a decisive advantage for domestic business.
2. Most permissive open-source license: both Coze Studio and Coze Loop are Apache 2.0 — commercially usable, privatizable, and customizable — making them better suited than Dify's non-OSI license for multi-tenant commercialization.
3. Widest L2 execution surface: the cloud PC (Ubuntu) + cloud phone (Android 13) expand an Agent's execution environment from API calls to a complete operating system.
4. L5 has a first-class citizen component: Coze Loop is open-sourced as an independent evaluation and operations component.
5. Strongest commercial validation: ARR $100M, 5–10 million messages consumed daily, 10,000+ registered enterprises — the quantitative evidence is the most complete.
6. Strong foundation: Volcano Engine 430+ trillion tokens called, 1 million+ enterprise customers, presence in 50+ industries.
7. Complete enterprise-side governance narrative: agile (ArkClaw) + steady (HiAgent) + unified governance (AgentSphere).

7.2. Weaknesses

1. L6 info missing on the platform side and open-source side: governance capability is concentrated in the paid enterprise offering; self-hosting Coze Studio requires building governance from scratch.
2. Conflicting ecosystem-data figures: plugins 100+ / 700+, stars 3k / 21.5k — sources differ widely and must be cited with caution.
3. L4 lacks engineered state management: there is no public run checkpoint, artifact versioning, or crash-resume mechanism.
4. Low predictability of multi-Agent dynamic routing: Agent World's flexibility and predictability inherently conflict.
5. Cost pressure is already apparent: 5–10 million messages consumed daily means a huge cost base for deep chains.
6. Strong lock-in to the ByteDance ecosystem: the value of multi-channel publishing depends heavily on whether Feishu / Douyin and other ByteDance channels are used.

7.3. Applicability Boundaries

7.4. Selection Recommendations

• Trade-off with Dify (both are Agent Platforms): choose Coze for channel distribution + permissive commercial license + open-source evaluation component; choose Dify for mature RAG depth + global community ecosystem + Human Input node.
• If a team is already in the Feishu / Douyin ecosystem, Coze's channel capability is enough to cover all other differentiators.
• If you plan multi-tenant commercialization based on the open-source code, Coze Studio's Apache 2.0 is clearly superior to Dify's Open Source License.
• If the primary need is governance and compliance, evaluate Volcano Engine's enterprise offering (HiAgent / AgentSphere) rather than self-hosting the open-source version.
• Before adoption, you must confirm: whether Coze Loop's evaluation capabilities meet your regression requirements; whether the open-source version provides the RBAC and auditing you need; and whether the cost of deep chains is within your budget guardrails.

Information Gap Statement

1. GitHub stars figures seriously conflict: the project hard data is "3k GitHub stars two months after open-sourcing", while a third-party repo stats site snapshot on 2026-09-06 shows 21,546 stars, and another official claim states "Coze Studio surpassed 10k within three days of open-sourcing, Coze Loop 3000+". The statistical subjects and time points of all three differ, and all are listed here without unification.
2. Latest version and feature availability: Coze Studio and Coze platform versions change rapidly with the release cadence; no authoritative version number was obtained as of the retrieval date 2026-09-12, marked [To be verified].
3. Plugin / Skills count figures: the platform built-in 100+ and third-party's claimed ecosystem 700+ coexist; the 365 Skills is the Coze 2.5 claim. All are from different sources, marked [To be verified].
4. Details of Coze Loop's evaluation capabilities: metric types, regression-set management, and trajectory-scoring methods were all not found in official documentation, marked [To be verified].
5. Coze 2.5 / 3.0 features: Agent World, persona / equipment / skill, Project Space, etc., come from third-party media reports and Volcano Engine public claims, and were not verified item-by-item against official release notes.
6. Quantitative data for industry cases: the deployment status in finance, education, retail, energy, etc., comes from IDC report retellings and Volcano Engine public claims, with no third-party independently audited quantitative effectiveness data, and no fabrication was added.
7. Pricing details: ArkClaw's lightweight version starting at CNY 29/month is a third-party claim; no official price list was found for HiAgent and the Coze platform's specific pricing, marked [To be verified].
8. Definition of "10,000+ Machine Learning related registered enterprises": the statistical definition of this hard data (industry classification, registered-entity type, time point) has not been further explained, marked [To be verified].
9. Platform-side L6 governance details: no public material was found on the implementation of RBAC, audit logs, data residency, and cost guardrails on the consumer / developer platform.

8. References

1. Coze Studio repository — coze-dev (GitHub). https://github.com/coze-dev/coze-studio
2. Coze Studio: All-in-One Visual AI Agent Development Platform — RepoRank (architecture, tech stack, star-trend snapshot). https://reporank.net/en/repo/coze-dev-coze-studio.html
3. 双第一！火山引擎领跑中国智能体开发平台市场 — 今日头条 (IDC report and industry deployment). https://m.toutiao.com/article/7651874887891468836
4. AI 智能体大战打响：字节、腾讯、百度都在卷什么 — 今日头条 (Coze 2.5, Agent World, ArkClaw / HiAgent). https://m.toutiao.com/article/7630374519244440074
5. 硅基员工时代来临：企业级 AI Agent 竞争版图深度解析 — 博客园 (Coze 3.0 upgrade highlights). https://www.cnblogs.com/haye-zhi-neng-guan/p/22789111
6. Apache-2.0、可商用、可私有化：字节把「扣子」核心引擎开源了 (open-sourcing time and license description). https://ima.qq.com/wiki/
7. 扣子 — AI 智习室 (platform positioning, plugin count, cumulative Agent count, and daily calls). https://aizxs.com/tool/coze-cn
8. Farewell to 'Card Stacking': 2026 Marks China's Computing Power Shift to 'Token Value Output' Era — AsiaICT (Doubao daily token growth curve, Volcano Engine MaaS share). http://www.asiaict.com/cloud/19273.html
9. Project parameter card v1.0 (Agent Platform definition and six-layer capability model, verified hard data) — internal baseline document of this project.
10. R01-Overview research report (concept boundaries and three-generation architecture evolution) — internal research report of this project.