Descriptions:
At the AI Dev 26 conference in San Francisco, AMD engineers Eda Zhou and Mahdi Ghodsi lead a hands-on workshop teaching attendees how to build personal AI agents using open-source models on AMD hardware, hosted by DeepLearningAI. The session walks through the full stack: deploying a large language model on an AMD Instinct GPU using ROCm (AMD’s open-source alternative to CUDA), connecting it to Open Claude — an open-source implementation of the Claude Code agentic framework — and building an agent capable of autonomous, multi-step task execution.
Ghodsi grounds the workshop in clear conceptual distinctions: a bare LLM is a stateless text generator; adding conversation history in the prompt creates a chatbot; true agents require the ReAct loop (Reason, Action, Observation) plus persistent memory, planning, and the ability to call external tools. Open Claude’s configuration system uses a set of markdown files — SOUL.md (behavioral guidelines), AGENTS.md (operating rules), and IDENTITY.md (persona) — that the agent reads and overwrites based on natural language prompts, allowing it to configure itself without manual file editing.
Attendees were given access to 120 dedicated AMD GPU instances to follow along in real time, with the practical goal of connecting an open-source model to MCP servers and real-world APIs. The session is a concrete demonstration of AMD’s ROCm ecosystem positioning as a viable alternative to Nvidia’s CUDA for AI agent development, with all featured models — including recent frontier releases — supporting day-zero AMD Instinct compatibility.
📺 Source: DeepLearningAI · Published May 20, 2026
🏷️ Format: Hands On Build
