Andrew Ng's agentic workflow patterns

Andrew Ng ran a series in The Batch this spring on what he calls agentic workflows, built around four design patterns: reflection, tool use, planning, and multi-agent collaboration. The core idea is that instead of asking a model for its final answer in one shot, you let it work in steps. It drafts, critiques its own output, uses tools, and improves before anything comes back to you.

I keep coming back to the series because it describes AI one step past where I actually am. My use is still one prompt at a time, with me reviewing everything that comes out. The reflection pattern is the one that keeps bugging me, in a good way. The review step is the part I've insisted stays human on my team all year. If a model can do some of its own checking before I ever see the output, I want to understand what my review is for and where it moves.

I don't have anywhere to apply this yet. It's the first AI writing in a while that made me feel like the ground is moving, though, so I'm keeping it close.

Link: https://www.deeplearning.ai/the-batch/agentic-design-patterns-part-2-reflection