From Microsoft and Google to Tencent Cloud, major tech giants have launched their own agent development frameworks and platforms in an attempt to gain a place in this field. This article will deeply analyze the latest trends in agent development in 2025, and discuss the background, current situation and future development direction of agent development entering the ecological stage.
In the past month, Microsoft build, Google I/O, and Tencent Cloud conferences have been held one after another, and today I would like to introduce to you, as of June 19, 2025, what users are suitable for the four major agent development frameworks (Agent Framework Development) and the eight major agent development platforms (Agent Development Platforms)? Which is the best to use?
1. Agent development method
Agent development is mainly divided into two categories: Agent Framework Development and Agent Development Platform. These two approaches represent different development paradigms, from highly customized at the source code level to convenient development on a platform.
Agent Framework Development
The core feature of agent framework development is source code development, which provides a “written agent architecture source code” and “all workflow node functions”, but the specific content needs to be written by the developer himself, and the workflow needs to be redesigned and arranged by the developer. This indicates that framework development pays more attention to underlying control and high customization, and developers need to be deeply involved in the design of the agent’s internal logic and workflow.
Features Overview:
- Source code-level control: Source code development is adopted, and developers directly operate the code.
- Architecture provision, content self-filling: The framework provides the architect architecture that has been built, but the specific business content needs to be implemented by the developer himself.
- Workflow rearrangement: Workflows need to be designed and orchestrated by developers themselves, fully reflecting the ability to customize.
Mainstream framework development tools:
1）OpenAI Agents SDK：
After 10 years of interaction design, why did I transfer to product manager?
After the real job transfer, I found that many jobs were still beyond my imagination. The work of a product manager is indeed more complicated. Theoretically, the work of a product manager includes all aspects of the product, from market research, user research, data analysis…
View details >
Functional positioning: OpenAI’s official open-source agent development framework provides lightweight code packages, allowing developers to build complex AI agents with minimal abstraction. The SDK aims to standardize agent communication and state management, making the development process more efficient.
Technical features:
- Python-first framework that supports building single-agent and multi-agent workflows;
- Built-in tool usage capabilities (such as network search, file retrieval, etc.) and monitoring and debugging functions facilitate the execution and observability of the agent’s decisions.
User-Oriented: Developers and enterprise teams across all technical levels who want to build their own AI Agents using the OpenAI platform.
Applicable scenarios: Typical applications such as customer service automation, multi-step research, content creation, code review, and lead follow-up, where complex tasks require AI agents to invoke tools autonomously and reason multiple rounds.
Open source time: Officially released on March 11, 2025.
2）Microsoft 365 Agents SDK：
Functional positioning: Microsoft’s agent development SDK for enterprises can build enterprise-grade, scalable, multi-channel AI agents with code. It is an evolution of the classic Bot Framework in the era of generative AI, blending conversation management, tool orchestration, and modern large model capabilities. Technical features:
- Support C# (JS and Python will be provided later) for development, and build agents that can be deployed in Microsoft Teams, Microsoft 365 Copilot, Web, and other channels.
- Integrate with any AI service (such as Azure AI Foundry/OpenAI) and orchestration layer (Semantic Kernel, etc.), and integrate with low-code agents created by Copilot Studio in both directions.
User-oriented: Enterprise developers who are deeply involved in the Microsoft ecosystem and want to customize and deploy internal AI assistants in a code-like manner.
Applicable scenarios: Intelligent assistants (such as office assistants, knowledge base Q&A bots), customer-facing conversational agents, and multi-agent process automation combined with business systems.
Open Source: Released in preview and open code on November 19, 2024.
3）Google Agent Development Kit(ADK)：
Functional Positioning: Google’s open-source agent development framework emphasizes flexibility and modularity for developing and deploying AI agents. ADK optimization adapts to Gemini models and Google Cloud services, but it is independent of the underlying model, deployment environment, and compatible with other frameworks. The original intention is to make agent development as efficient and controllable as regular software development.
Technical features:
- Provide multiple built-in workflow agent types (sequential, parallel, loop) and LLM-driven dynamic routing to achieve predictable and adaptable process orchestration;
- Support for a wide range of model access (through Vertex AI Model Garden and LiteLLM, seamless integration of multiple models such as Anthropic and Meta) and a rich tool ecosystem (built-in function tools, integration of LangChain, LlamaIndex, and even other agents as tools);
- It supports two-way audio/video streaming interaction for more natural multimodal dialogue.
- Provide local command-line and visual web interface debugging to view status, track execution steps, and debug agent decisions, with built-in evaluation tools for testing agent performance and behavior.
User-oriented: Professional developers of the Google Cloud ecosystem and development teams who want to use open source frameworks to independently control complex multi-agent applications.
Applicable scenarios: Complex applications that require strict control over agent reasoning processes and collaboration methods, such as assistants involving multimodal interactions such as voice dialogue, AI services with multi-model integration, and enterprise projects with independent and controllable requirements for the entire development and deployment process.
Open source time: In April 2025, open source was announced at the Google Cloud Next 2025 conference.
Agent Development Platform
Platform Type:
- Codebase: A platform that provides an agent framework or code snippet.
- Low-code: Allows development with a visual interface and a small amount of code, including visual components.
- No-code: Agent development through conversations with AI assistants (essentially the agent itself is a program).
At present, the main international popular platforms:
- LangChain/LangGraph (foreign users): Very popular in the field of agent development, often used to build complex chain workflows.
- Dify (Domestic Users): A platform that provides a one-stop integrated development environment.
- n8n (International Users): Versatile automation and workflow orchestration tools that can also be combined with agent capabilities.
- Coze (domestic users): An emerging platform positioned as an online low-code Q&A assistant agent platform.
1）Langchain/LangGraph
Functional positioning: open source code base + visual extension; High flexibility
Technical features: Full ecosystem: The LangChain ecosystem includes LangStudio (development platform with visual interface), LangSmith (cloud agent operation and maintenance platform, similar to Docker), LangGraph (dedicated agent development Python library), etc., with rich and complete community components.
User-oriented: Enterprise users who need any in-depth customization of agent applications can be privatized and deployed.
Advantages: The most complete function and the most prosperous ecology. (LangChain, which has been focusing on agent construction since October 2022 and has a user base in the field of strengthening learning; Highly modular, with a high community maturity, solutions can be found in almost all enterprise scenarios in the LangChain community.
Disadvantages: high complexity and high learning costs.
2）Dify
Functional positioning: open source low-code platform; One-stop development; Integrate back-end as a service (BaaS) philosophy with LLMOps practices.
Technical Features: Low Threshold Feature-rich: Dify stands out by supporting both low-code operations and engineering deployment needs, finding a balance between “business” and “technology.”
User-oriented: Quickly implement AI applications within the enterprise and support privatization deployment.
Advantages: Quickly build and implement Agent + RAG + Workflow from scratch, and the learning cost is lower than that of LangChain.
Disadvantages: Haute Couture is limited by official support. (For example, I thought “twitter” but dify doesn’t, github does, and you need to extend it yourself.) ）
3）N8n
Feature targeting: Visualize the workflow orchestration engine
Technical features: AI workflow automation: n8n, formerly the open-source version of Zapier, supports 400+ integrated nodes for different services and applications, and now introduces LLMs to upgrade automated processes to AI workflow automation.
User-oriented: Enhance existing business process automation (RPA + AI) with AI, support private deployment, and be ideal for content creation and other scenarios.
Benefits: Widest application integrations (over 422 app plugins to connect to almost any website; visual orchestration is easy to use).
Disadvantages: Non-dedicated AI agent development platform (mainly workflow orchestration), advanced agent functions need to be implemented additionally, and the ecosystem is relatively imperfect.
4) Coze (Button)
Functional positioning: online low-code “Q&A assistant” agent platform (positioning Q&A assistant agent)
Technical features: Out of the box, extremely low threshold: Coze is one of the earliest agent development platforms in China, integrating 60+ domestic plug-ins such as WeChat public account, AutoNavi Map, Feishu, Baidu, etc.
User-oriented: novice/individual user, the lowest learning cost; It is suitable for scenarios such as conversation flow and allows for private deployment (enterprise users).
Advantages: Built-in 60+ local plugins; Learning costs are lower than Dify.
Disadvantages: The model is mainly in-house (Volcano Engine is excellent in content creation, but poor in other aspects, such as code, research and analysis), lacks components in multi-agent construction, and has limited flexibility (mainly because Chatflow’s agents are not collaborative).
Emerging agent development platforms:
Tencent Cloud Agent Development Platform (and Tencent Components): The solutions provided by large domestic cloud service providers help enterprises activate private knowledge and quickly build exclusive agents, indicating that enterprise-level agent development platforms are rising.
Tencent Yuanqi: https://yuanqi.tencent.com/
Tencent Cloud Agent Development Platform: https://lke.cloud.tencent.com/
Microsoft Copilot Studio: Low-code/no-code development platform with built-in Copilot Q&A assistant, allowing agent development through natural language; Built agents can be deployed to the Microsoft cloud for free for use with Teams and Microsoft 365 Copilot.
Copilot Studio: Access directly from the Copilot assistant
Website: https://copilotstudio.microsoft.com/
Azure AI Foundry: An enterprise-level agent development platform (also a general-purpose AI application development platform).
advantageBoth Azure and Microsoft Copilot Studio support MCP and A2A protocols (meaning agents developed with the OpenAI Agents SDK framework can communicate collaboratively with agents built on platforms such as Dify and LangGraph through the Microsoft agent platform).
Website: https://ai.azure.com/
Google Vertex AI（Generative AI App Builder）: Google Cloud’s agent development module, a one-stop building platform for enterprise AI projects.
Website: https://console.cloud.google.com/
2. Trends in agent development platforms – simpler, more efficient, and more integrated
The germ of open standards (e.g., MCP, A2A protocol) indicates that agents of different frameworks may be able to communicate and collaborate with each other in the future—just like computers from different manufacturers connected to the same Internet.
When this interconnection is realized, the multi-agent ecosystem will truly form a network effect, and the boundaries of agents’ capabilities will be greatly expanded.
Various intelligent development platforms are working hard to lower the threshold and accelerate the implementation.
From enterprise-level Azure Foundry and Tencent Cloud platform to popular Copilot Studio, Coze, and Yuanqi, they serve different user groups with different entry points, but the trend is simpler, more efficient, and more integrated.
The platform considers many things in advance for users: common tools are directly provided, industry processes are directly built-in, deployment and maintenance are directly managed, and security and compliance are directly escorted.
This allows non-AI professionals to participate in the creation of AI applications and unleash their imaginations. It is foreseeable that in the future, product managers and business experts will co-create AI agents with programmers, just as it is common for “citizen developers” to use low-code tools today.
The evolution of large models themselves also affects the direction of agent development: the stronger the model ability, the more “smart” the agent will be and can complete more complex tasks autonomously. But it also poses reliability and security challenges, which is why frameworks and platforms emphasize guardrail and content security. How to give agents greater autonomy while ensuring their control and trust will be an ongoing issue.
3. Market status and judgment
At present, the field of agent development is showing a diversified competition situation: OpenAI, Microsoft, Google, etc. have successively opened up agent frameworks internationally, eight mainstream platforms are blooming, and domestic manufacturers such as Tencent are also accelerating their layout.
Each solution has its own focus: framework development gives developers a high degree of flexibility and control, while platform development greatly lowers the threshold for use and speeds up application implementation.
Overall, industry standards have not yet been unified, but open protocols such as MCP proposed by Anthropic and A2A promoted by Microsoft are promoting the interconnection of the industry ecosystem.
The agent development market is still in a stage of rapid evolution, and there is no absolute optimal solution, and users of different scales and technical strength can choose the most suitable path according to their own needs.
4. Potential opportunity direction
Under the pattern of a hundred flowers blooming, there are still many subdivision opportunities worth paying attention to.
First of all, vertical industry solutions: Customized agents for specific fields such as finance, medical care, and customer service, combined with industry knowledge and compliance requirements, to provide in-depth functions not covered by general platforms.
secondly, cross-platform integration and management: Develop middleware or management tools that connect different frameworks and existing systems of enterprises (using open protocols) to achieve unified orchestration and monitoring of heterogeneous agents, and solve the pain point of “multiple agents fighting independently”.
furthermore, low-threshold creation and ecology: Build simplified tools that allow business personnel to participate in customizing agents (such as template libraries and plug-in markets), lowering the threshold for use while forming their own ecological stickiness.
5. Product strategy suggestions
In the face of this situation, product strategies should leverage strengths and avoid weaknesses.
On the one hand, clarify its own positioning, and choose the ecological position according to the target users: it can be deeply integrated into the Microsoft or Google system, with the help of its mature infrastructure and customer base; or take the open source and independent route to differentiate breakthroughs in the local market.
On the other handAdhere to openness and compatibility, and avoid working behind closed doors: Priority support for general protocols such as MCP and A2A, compatible with multiple models and platforms, and prevent over-reliance on a single vendor leading to being controlled by others.
At last, focusing on user value and experience: providing easy-to-use and efficient development interfaces and pre-component modules, so that non-technical users can also participate in building agents; At the same time, it has built-in complete security measures and operational tools to improve the reliability and controllability of the agent and win the trust of enterprises.
In the context of “framework vs. platform”, flexibly integrating the two paradigms and quickly delivering actual business value is the way for agent products to win.
