CMU AI Agents 2026: 1. What are Agents and How Do They Work?

AI agents function as systems that perceive and act upon environments, ranging from code repositories to web browsers. Effective agent development requires balancing "harness engineering"—the creation of structured systems for memory, tool use, and safety—with direct language model training via reinforcement learning and supervised fine-tuning. While current agents demonstrate significant potential in tasks like automated coding and GUI navigation, they remain prone to errors, such as context loss or unsafe actions, necessitating robust sandboxing and monitoring. The course emphasizes mastering these technical trade-offs by implementing agents from scratch, designing evaluation frameworks for multi-step tasks, and applying reinforcement learning to improve agentic performance. Students will gain hands-on experience by building and refining these systems, ultimately moving beyond simple prompt engineering to develop reliable, autonomous agents capable of complex, long-horizon reasoning.