Summary of Key Points
Over the past three years, Coze has completed two significant evolutionary steps: first, it has upgraded its development approach from a low-code model based on drag-and-drop building blocks to natural language programming, where users simply state their requirements and the AI takes care of the rest; second, it has evolved its collaboration method from individual AI agents working independently to teams composed of multiple AI agents collaborating together. The latest version 3.0 marks an important milestone in this collaborative evolution, addressing issues such as context loss for single AI agents, device synchronization issues, and limited capabilities. The article also compares the development paths of Coze with platforms like Dify and n8n, highlighting that the future competition among such AI tools will shift from the strength of their features to the accumulation of industry-specific knowledge (KnowHow). After all, while AI can execute tasks, the decision on what to do and to what standards to adhere remains in the hands of humans.
Detailed Breakdown
1. Coze’s Two Evolutionary Paths: Working Together Like Two Legs
Coze’s improvement has not been a linear process; instead, it has progressed along two parallel paths:
- Development Path: Focusing on making building AI applications simpler. Version 1.0 used drag-and-drop nodes (similar to constructing with Lego, connecting functional modules); Version 2.0 introduced natural language commands for specifying tasks directly, allowing the AI to generate the necessary processes automatically.
- Collaboration Path: Focusing on improving efficiency in task completion. Both versions 1.0 and 2.0 relied on a single AI agent handling all tasks; Version 3.0, however, involves multiple AI agents working in teams (e.g., one AI for data research, another for writing reports, and yet another for polishing the text).
These two paths have intersected, transforming Coze from a low-code tool into an AI team collaboration platform.
2. The Transition in Development Methods: Why Stop Using Building Blocks?
Coze 1.0’s drag-and-drop interface was user-friendly for beginners, but it became cumbersome for more complex tasks. For instance, setting up a workflow with hundreds of nodes resembled weaving through a spiderweb, and making even minor changes could have significant impacts on the entire system, making debugging extremely difficult.
As AI capabilities improved significantly by 2025 (with a substantial leap in AI programming skills), users no longer wanted to manually drag and drop nodes. Coupled with the emergence of AI-based programming tools like Cursor, Coze had to shift to natural language programming. Now, users simply need to clearly articulate their requirements (e.g., “Help HR filter resumes and store them in a Lark spreadsheet”), and the AI will generate the necessary processes automatically. This increase in efficiency comes with new challenges, such as potential misunderstandings by the AI and the need to review code logs during debugging, which can be frustrating for inexperienced users.
3. The Evolution of Collaboration: From Single AI to Teamwork
Single AI agents face several major drawbacks:
- Limited Memory for Long-Term Tasks: They may forget previous progress on tasks like weekly reports.
- Device Isolation: Work completed on one device might not be accessible on another (e.g., on a phone but not on a computer).
- Limited Capabilities: An AI agent cannot simultaneously perform multiple tasks, such as market research, content writing, and creating PowerPoint presentations.
Coze 3.0 addresses these issues with a “project + multi-agent team” approach. Users define a goal (e.g., “Create a report on the coffee industry for 2025”), and the platform automatically divides the tasks among different AI agents (researching data, analyzing results, creating visuals). This ensures seamless progress across devices (phone, computer, web), allowing users to focus on setting directions and reviewing outcomes without worrying about the intermediate steps.
4. Divergent Platform Paths: Each with Its Target Audience
In response to the trend of AI programming, different platforms have chosen different approaches:
- Coze: Completely abandoned drag-and-drop functionality, targeting beginner users who do not want to see complex nodes.
- Dify: Still uses visual nodes, catering to enterprise users who need control over the process (e.g., understanding how each step is executed), but its community allows for AI-generated code imports.
- n8n: Retains nodes but uses AI to assist in building the workflow, appealing to technical enthusiasts who prefer flexibility without manual drags and drops.
The differences reflect the varying user groups each platform serves, but all aim to reduce manual operations and leverage AI to streamline tasks.
5. The Future Focus of Competition: Not Tools, but Industry Knowledge
The article emphasizes that while AI tools will continue to improve in functionality (all will enable team-based collaboration), the true differentiator will be industry-specific knowledge and expertise. For example, while AI can help with resume screening, the criteria for evaluating candidates (e.g., 5 years of experience, background in a large company, Python skills) rely on human expertise. Similarly, although AI can analyze financial data, the logic behind making investment decisions (e.g., using PE ratios or understanding industry trends) also comes from human knowledge.
These industry-specific aspects are irreplaceable by AI and will become the key differentiating factors among future platforms.
Final Conclusion
AI tools will continue to advance, but ultimately, it will be human knowledge—specifically, the ability to determine what is needed and to set clear standards for task completion—that will determine their effectiveness. This is what remains most scarce and valuable.