This is Andrew Ng’s short talk “What’s Next for AI Agentic Workflows,” given at Sequoia Capital’s AI Ascent event and posted on the Sequoia Capital YouTube channel in March 2024. Ng co-founded Google Brain and Coursera and runs DeepLearning.AI and AI Fund, and he uses the roughly fourteen minutes to argue that how a model is used can matter as much as which model it is.
His central point is that agentic workflows, in which a model iterates rather than answering in a single pass, can substantially raise quality. He groups the useful patterns into four families: reflection, where the model critiques and revises its own output; tool use, where it calls external functions; planning, where it breaks a task into steps; and multi-agent collaboration, where several model instances cooperate. He cites results in which an agentic workflow built on an older model outperformed a stronger model used in one shot.
For a business reader, this is a compact, firsthand framing of why “agents” became the dominant theme in applied AI. It explains, without hype, why wrapping a model in an iterative loop is often the cheapest path to better results.