Stanford University's CS329A, titled "Self-Improving AI Agents," is a graduate-level course available for free on YouTube, taught by industry experts Azalia Mirhoseini and Okachandra Chowdhery. The curriculum focuses on agents that achieve continuous improvement through environmental interaction and self-feedback rather than static updates. A core technical pillar of the course is the concept of test-time compute, where increasing a model's reasoning time during inference can be as effective as scaling parameters. The 20-week syllabus covers critical topics including robust verification, reinforcement learning (RL) using methods like STaR and DAPO, and multi-step reasoning via tree-search. A significant technical takeaway emphasized is that the primary bottleneck in agent development is the verifier (the mechanism used to validate correct answers) rather than the generator itself. The course highlights how high-quality, self-generated data from successful reasoning loops can be used to fine-tune subsequent model generations. Key methodologies discussed include Mixture-of-Experts (MoE) architectures, reinforcement learning from human feedback (RLHF), and agentic patterns such as parallel execution and centralized orchestrators. The course is designed around real-world research, incorporating seminal papers like ReAct and recent advancements in long-running agent benchmarks like METR.