Why Agentic AI Works when LLMs Fail

After several conversations with friends (you know who you are!), I’ve decided to shift the focus of this post. Rather than diving into the different platforms for implementing Agentic AI, I want to first explain what Agentic AI is and share my thoughts on why it works.

Agentic AI by Analogy

A simple way to think about it..? People build knowledge systems where the output is greater than any one individual can provide.? Consider the software teams you have encountered in the past.? In a high-functioning team, everybody contributes something unique to the project that nobody else considered.

So why not think of AI in an analogous way?? Can a bunch of mostly reliable, and yet clearly imperfect entities collaborate to produce a superior result?? Yes, they can.

Dr. Andrew Ng is a pretty smart guy who founded Coursera and Sequoia’s AI Fund, and he is an adjunct Stanford professor to boot.? Andrew and his Stanford students showed that a team of AI agents using ChatGPT 3.5 can outperform a single instance of ChatGPT 4.0, as shown in the chart.? It is a bit to interpret, but the simple takeaway as that multiagent ChatGPT 3.5 always outperforms a single ChatGPT 4.0 interation. I use ChatGPT in this example because of its familiarity, but the trend holds when using different LLMs as well.? I highly recommend checking out his 15 minute primer on Agentic AI.

What is Agentic AI?

Rather than engaging in a single question / answer session with a chatbot as with ChatGPT, Agentic AI works by creating a team of agents to solve a problem, and those agents are specialized and can collaborate with each other.? In general, the agentic workflow is:

1. Develop a step-by-step plan for how to solve the problem.
2. “Hire”? (e.g. create) a team consisting of team members with the rights to implement each step.
3. Have the team collaborate to create the solution. Each agent listens for input from others, does its part, and shares the results with the team.

Upon reflection, you may still be wondering why you need multiple agents.? Why not give all the necessary skills to one agent and say “go solve this problem”?? The answer lies in collaboration.

Consider what happens on a human software development team consisting of a product manager, a developer and a test engineer (sounds like the start of a terrible joke, I know!).? Both the engineer and the tester receive the same requirements from the PM.? The engineer implements code to meet those requirements, passes the code to the tester, who finds that some of the test cases fail.? Many test failures require some discussion between the developer, tester, and PM in order to sort out the intended behavior.? On highly performant teams, these discussions lead to a better, more reliable product for users.

In almost exactly the same way, an Agentic AI-based software team would see the same dynamic.? The tester and developer agents can clarify the requirements with one another, and involve the PM agent if they cannot agree.? Once clarified, the appropriate code and documentation can be updated and the team can iterate to achieve the aligned requirements.

In the algorithm above, all of the iteration required to solve the problem happens in step 3, but is all determined on demand.? So the path isn’t really deterministic.? In fact, two executions by the same Agentic AI team will most likely produce different execution paths to solve the same problem.

I am leaving out a ton of details about broader Agentic AI in order to fit the medium, but I am highlighting multi-agent collaboration because that is the core idea that changes everything.??

Why Does It Work?

More accurately, this title should be “why does Glenn think it works?”? I am not an academic researcher and while I have an intense intellectual curiosity, I am unapologetically motivated to feed my family.? So I’m offering a rationale that I believe is logical, though I have absolutely no proof - now or in the future.? Any academics out there should feel free to dig in.

It works because of the distinct “personalities” of each agent. Agents can have different skills, tools, and even response styles. These variations impact how they interpret and handle tasks, creating gaps where ambiguity exist, and lead to creative problem-solving that identifies and resolves ambiguity well. In human teams, it’s often the dialogue between different perspectives that leads to breakthrough solutions. I believe the same is true with Agentic AI.

That’s my theory, and I’m sticking to it - at least until some academic proves me wrong!

Short Advert for Me

This year has been one for the books. After growing my business 10x in two years, I was laid off last December as part of a consolidation. Just as I was about to jump back into the job market, my mom was diagnosed with a terminal illness. I spent the next six months caring for her. During that time, I also took a remote AI class to formalize and expand my knowledge. Two weeks after my mom passed, my son defended his PhD thesis in Materials Engineering (huge congrats to Matthew Chagnot !). Now, I’m back in the job market, and while it’s a tougher landscape, I wouldn’t change a thing. I’m incredibly grateful for the time I had with my mom.

If you are an employer, my posts focus on Agentic AI because I find it fascinating and relatable to many. Please know that I bring a wealth of AI-based techniques to my core expertise in networking and cloud. If you're looking for someone with practical AI knowledge and deep experience in cloud and networking, let’s connect!

Your Glad it isn't over 100 in Texas Anymore Friend,

Glenn