The video guides viewers through building intelligent agents on the AMD Developer Cloud using MI300 GPUs, demonstrating how to set up VLM inference, create OpenAI-compatible agents with PyAndic, and enhance them with custom tools and MCP servers for real-time data access. It showcases practical applications like retrieving the current date and performing Airbnb searches, highlighting the platform’s capabilities and encouraging users to explore agent development with available free credits.
The video provides a step-by-step guide on how to build your own intelligent agent using the AMD Developer Cloud. It begins by instructing viewers to visit devcloud.amd.com, where they can spin up a GPU droplet equipped with MI300 GPUs. Users are then advised to select the VLM quick start package to get started with VLM inference and add or select their SSH keys before creating the droplet. After a brief wait for the droplet to initialize, users can connect via SSH or access the Jupyter notebook using a pass key obtained from the web console.
Once connected, the video demonstrates how to verify the GPU status by running the “rocm-smi” command, confirming the MI300X GPU is active. The next step involves starting the VLM server through a command in the notebook, with GPU utilization monitored to ensure the model is loading correctly. A test command using curl confirms that the open-source Qwen 30B model is running successfully on the VLM server, setting the stage for building the agent.
The video then explains how to create an agent using PyAndic to make an OpenAI-compatible endpoint and instantiate the agent. An asynchronous agent function is defined to test the setup, with a simple example confirming that the model correctly answers factual questions like the capital of France. However, the video highlights a limitation of large language models (LLMs): they rely on their training data and may not know current facts, such as today’s date, which can lead to hallucinations or refusals to answer.
To address this, the video shows how to create a custom tool that provides the current date to the agent. This tool is integrated into the agent, enabling it to answer questions about today’s date accurately. The presenter then introduces MCP (Model Context Protocol) servers, which are community-built tools that can be leveraged by agents. By replacing the custom date function with an MCP time server, the agent can access the current date in a more standardized and shareable way.
Finally, the video explores a more complex example using an MCP server for Airbnb search to help plan trips. After installing additional dependencies and updating the agent with the Airbnb search tool, the agent can search for accommodations in a specified location and date range. The agent retrieves options and provides links to Airbnb listings, demonstrating practical real-world applications. The video concludes by encouraging viewers to try building their own agents on the AMD Developer Cloud, with free credits available to get started.