How to Develop Intuition About AI Agents

Link to slides

I have spent a large part of my time (from 2023-2025) thinking about AI agents (yea, before they became the buzzword of 2025). When I joined an edtech company as the first engineer in the brand new Agentic AI team, I built a tiny agentic AI framework, and made a company-wide demo showing a turn-based interaction, wherein agents held debates and an umpire judged winners based on a pre-configured rubric.

Toy example, but the point was made: given the quality of the LLM tech at the time, agentic AI was inevitable and potentially worth exploring.

Soon afterwards, I stopped maintaining my framework and started using the nascent Microsoft Autogen framework, and it was with this that I shipped the v0 of our first agentic AI product.

Being that early comes with its own battlescars, and I may still be recovering from burnout from the experience. Still, I think of my work fondly. It forced me to grapple with things I’d otherwise never think about in a more mature ecosystem. Things like agent orchestration, messaging protocols, downstream quality degradation (a poor-performing agent early in a multi-agent system will necessarily downgrade the output quality of agents that depend on its outputs), structured outputs (less of a problem today, but not a truly solved problem as there still isn’t a ‘standard’ way of dealing with these things today).

Then there are the evals, tests and prompt optimization. The bigger the head, the bigger the headache, and multi-agent systems are truly multi-head problems.

I felt somewhat vindicated when Anthropic reported on the state of the industry in 2024. Many people who were building agentic systems were really building workflows, which were easier to reason about, as they were a specific kind of graph. Far from simple, but the simpler kind of graph.

I’d become interested in graph theory while working at TradeCore in 2021, because it was visual math, and was a great way to mentally model the relationship between things. It was also an easy introduction to pure mathematics, and I dabble. I’d read the first few pages of Introduction to Graph Theory, by Richard Trudeau (great book, absolutely recommend).

Seeing the vertices/nodes and edges in workflow builders made me realize that graph theory, while not entirely necessary for understanding multi-agent systems, presented a great abstraction for thinking about agentic AI systems. Since I was motivated to remove myself from the tyranny of young, fledgling agentic AI frameworks, I started to spend more time exploring the raw concept of modeling agent relationships as graphs.

Fun fact: the concept by which neural networks update their weights and biases during training (backpropagation) is also graph-based. There are graphs everywhere for those with the eyes to see.

This is my first ever public technical talk (I make this distinction from internal company demos). Thank you to the team at SysConf for inviting me. It is one of the highlights of my year.

Have a great weekend!