Stanford CS329A Self-Improving AI Agents | Part 8 | Agentic Evaluations and Long Horizon Tasks

Want to dive deeper? This curriculum is covered in the following online courses: - Agentic AI professional education course: https://stanford.io/4zOyjPN - XCS329 graduate course: https://online.stanford.edu/courses/cs329a-self-improving-ai-agents A similar curriculum is covered in XCS329z https://online.stanford.edu/courses/cs329z-engineering-ai-agents Follow along with the course schedule and syllabus: https://cs329a.stanford.edu/ View the course playlist: https://www.youtube.com/playlist?list=PLangBM27OtEA Video Summary: This lecture video from Stanford's CS329A, Self-Improving AI Agents, taught by Aakanksha Chowdhery on November 17, 2025, covers methods for evaluating AI agents on long-horizon and economically valuable tasks. It reviews METR's time-horizon methodology, which measures the task duration models complete at 50 and 80 percent reliability across the HCAST, SWE-bench, and RE-Bench suites, showing that model capability has roughly doubled every seven months, from seconds for GPT-2 in 2019 to nearly an hour for Claude 3.7 Sonnet in 2025. It also covers GDPval, OpenAI's benchmark that scores model output against work from industry professionals across 44 occupations and nine sectors, where win rates rose from 12.4 percent for GPT-4o to 47.6 percent for Claude Opus 4.1. Stanford's Deep Scholar Bench, also discussed, tests whether models can generate literature review sections for academic papers by scoring knowledge synthesis, retrieval quality, and citation verifiability. The lecture closes by detailing recurring agent failure modes, including poor planning, incorrect tool selection, premature task abandonment, and repetitive action loops. Speaker Bio: Aakanksha Chowdhery Adjunct Professor of Computer Science, Stanford University Dr. Aakanksha Chowdhery is pushing the frontier of agentic LLMs, focusing on recursive self-improvement and long-horizon agents that learn and deploy in the real world. She is one of the few researchers globally who has led frontier model training end-to-end, across both dense and mixture-of-experts (MoE) architectures. At Google, she led the 540B PaLM model, the largest densely trained language model in the world at the time. She subsequently drove pre-training and scaling of Gemini's MoE models across multiple generations, and contributed key components to PaLM-E, Med-PaLM, and the Pathways infrastructure underpinning Google's large-model efforts. She went on to build and lead pretraining teams for open intelligence efforts at Reflection and Meta. Earlier, she held research roles at Microsoft Research and Princeton. At Stanford, where she earned her PhD, she teaches CS329A (Self-Improving AI Agents) and serves as Program Chair for MLSys 2026.