Programming with GitHub Copilot Agent Mode

Microsoft

Apr 04, 2025

GitHub Copilot continues to grow, starting with code completion/generation/optimization, to GitHub Copilot Chat, which optimizes the quality of AI-generated code through chat interaction, to GitHub Copilot Edit Model, which can collaboratively create/debug/adjust based on multi-file associations in project engineering, and finally to today's GitHub Copilot Agent Mode, which completes project engineering through AI Agents. GitHub Copilot combines LLM to create an intelligent programming solution based on AI Agents, allowing non-computer practitioners to truly participate in project development. This blog will focus on GitHub Copilot Agent Mode to give you a deeper understanding of this intelligent programming method.

Be a game Developer

I believe everyone can play video games, but if you want to create a video game personally, there will be a certain threshold. Let's use GitHub Copilot Agent Mode to create your first online game.

Select Agent Model through GitHub Copilot, select Claude 3.7 Sonnet, and enter the following in the chat

Please help me generate an HTML5 basketball shooting mobile game according to the following requirements 1. All elements of the game are generated with svg 2. Players need to adjust the strength to complete the shot on the service line 3. The basket is directly above 4. There are 3 levels, and you need to complete the shot within a limited time. Each time you shoot 10 shots, and you can advance to the next level after you make 8 shots. The time for each level is decreasing, such as 30 seconds for the first level, 20 seconds for the second level, and 10 seconds for the third level 5. Each player needs to enter his name before entering the game 6. You can play again if you fail to pass the level

We can generate a basketball shooting game based on HTML 5 in 3-5 minutes

Try more models in GitHub Copilot Agent Mode

GitHub Copilot provides enough world-class programming models for users to choose from, such as Claude 3.5 / 3.7, GPT 4.5 / 4o, Gemini 1.5 and other models. For different development scenarios, especially in enterprise scenarios or models that require localized Ollama or third-party interfaces, such as OpenRouter, Azure OpenAI Service, OpenAI, etc. GitHub Copilot now supports more model access. We can add more third-party models through Manage Models when selecting a model.

After clicking, you can select models from different sources

Here we choose OpenRouter (if you don’t have OpenRouter, you can register to get the API Key at https://openrouter.ai/). If it is the first time to use it, you need to enter the API KEY, otherwise the model list of OpenRouter will appear, as shown in the figure:

We can try to use the latest DeepSeek V3 0324 model as the model for GitHub Copilot Agent Mode.

Next, we try to use

Create a private task management system through Python Flask. The frontend is saved in the templates folder. It uses HTML5 + JS + CSS. The backend uses JSON to save data.

MCP expands GitHub Copilot Agent Mode enterprise project capabilities

In enterprise programming, AI is not only needed to generate code, but also to connect to different tools, such as databases, cloud services, DevOps management, etc. We can connect to different tools through the MCP protocol and combine LLM to complete stronger enterprise scenario applications.

GitHub Copilot Agent Mode supports MCP servers, and you can connect to different MCP services. We can think of an interesting scenario, which is to make a project report ppt. In the past, you needed to invest manpower and time to complete it, but through GitHub Agent Mode and MCP, you can quickly complete this task. Let me explain it next.

We need a PPT MCP service. At this time, we can find an https://github.com/microsoft/semanticworkbench/tree/main/mcp-servers/mcp-server-office Office MCP service, which can start the SSE connection through local STDIO and remote SSE by executing the following command on the command line

uv run -m mcp_server.start --transport sse --port 25566

After successful startup, enter GitHub Copilot Agent Mode to configure MCP, as shown in the figure

Open the configuration file and add mcp server

"mcp-server-office": { 
    "url": "http://127.0.0.1:25566/sse", 
    "type": "sse" 
}

We try to download Semantic Kernel, and then use DeepSeek V3 0324 to analyze the project and generate a 10-page PPT with the following prompts:

Please analyze the project and make a 10-page ppt

In Agent Mode, we can analyze the Semantic Kernel project through DeepSeek V3 0324 and call PPT MCP Server to complete simple PPT content creation. This is just a small example, of course, you can also make better use of MCP to create more enterprise-oriented applications.