AI Agents and Agentic Workflows as the next evolution of AI

Great talk about the next step in AI, which is agentic workflows, by Andrew Ng at Sequoia Capital.

Instead of one shot prompting to one chatbot with prompt engineering, there is N shot prompting to any number of interacting AI Agents. What’s really cool is that lower capability LLMs can actually beat out higher capability LLMs. It’s the AI version of scaling “horizontally” rather than “vertically”.

So it’s really an optimization between the tech vs. capital costs, for example, to what extent do we make 1 API call to an expensive LLM like GPT5 or GPT6 vs. 100 API calls to very cheap open source LLMs to coordinate 50 agent. Many variations on this theme, but the general idea may be that fast inference (such as groq.com) of open source models may be far superior to major frontier models.

Here are the screenshots from the talk.

Inferior LLM models underlying agents perform better than superior LLM models used in zero shot prompting

Design patterns include reflection, tool use, planning, and multi-agent collaboration

Design Pattern 1: Reflection