TL;DR: Given the current price points of LLMs, you should always design your AI workflows to incorporate multiple agents, to refine, QC, and build on each other’s outputs.
I recently watched a fascinating video by Andrew Ng. It’s worth it.
The core message is—don’t look at AIs as singular tools that must fulfill a process through a single step. Consider AIs as fallible, imperfect agents that interact with each other to improve their capabilities and yield better results.
You know, a bit like teams of humans. After all, you wouldn’t forward an intern’s competitive landscape analysis to your CEO, right?
Let’s see what this means…
Here is a real-life example. I used Gemini for this purpose, but the same logic applies to any LLM.
Prompt: You are a high school teacher. Explain to a student what turbulent flow is, in a short paragraph.
Answer: Imagine a river instead of flowing smoothly like on a calm day, it swirls and eddies in all directions. That’s turbulent flow! It’s when a fluid, like water or air, gets chaotic and mixes itself up instead of flowing in neat layers. This happens because the fluid’s motion gets too strong and overcomes the “stickiness” that normally keeps things smooth. Smoke rising from a campfire or the wind whipping through trees are both good examples of turbulence.
Ok, not too shabby, besides the grammar of the first sentence. I’m a chemical engineer (well, I used to be, until my experience in a chemical lab, but that’s a story for another day), and this is reasonably accurate.
Now, let’s ask another AI agent to improve on it.
Prompt: You are a high school supervisor. How would you improve this definition of “turbulence” that one of your teachers gave?
Answer: Imagine a river. Instead of flowing smoothly like on a calm day, it swirls and eddies in all directions. That’s turbulent flow! It happens when a fluid, like water or air, gets chaotic and mixes itself up instead of flowing in neat layers. This happens because the fluid’s motion gets too strong and overcomes its viscosity, the resistance that normally keeps things smooth. Smoke rising from a campfire or the wind whipping through trees are both good examples of turbulence.
That’s better. There are no grammar issues, and it introduces the concept of viscosity.
This is the core concept behind an agentic workflow. Instead of having an LLM do all the heavy lifting in one pass, you can—and should—employ a team of LLM agents, each revising, critiquing, and refining the outputs of the previous step.
You can even bring it further.
Prompt: You are a high school supervisor. How would you rank the following text for clarity? Please give it a score from 1 to 10.
The answer is 9/10, with a recommendation to discuss the Reynolds number and its role in predicting turbulence, for an advanced audience. This would be helpful information for a human overseer.
So what should I do?
-
Think in terms of pipelines. When you interact with AIs and AI agents, think in terms of pipelines. For example, have an agent summarize a text, another one critique the summary for accuracy and conciseness, and a third one integrate the changes.
-
Set the right personas. Always make sure to set the right personas. “You are a competitive intelligence professional; summarize this text”; “You are a business executive: what critiques do you have about this text”
-
Use quality filters. Think, when appropriate, of quality filters. Create iterative workflows, where agents give feedback to each other—while regularly checking for accuracy—until quality metrics are met.
Originally published on LinkedIn in April 2024.