AI Agents may be the biggest craze in AI today. There is some good reason for this, of course, but most people have not thought critically about exactly what AI agents are, what they do, and how to make them work pragmatically. Here, I will lay out why I think very simple, specialized agents will be the norm. Those simple agents will be combined, however, to enable very complex and compelling functionality.
Complex “Agents” Are Really Lots Of Simple Agents Combined
My main point this month is that many of the “agents” you hear about aren’t a single agent. While “agent” is used to represent a set of activities that will be completed on your behalf, under the hood there may be many separate agents handling different parts of the task. This is a good thing because it means that the highly granular agents under the hood can be kept simple while still enabling very complex capabilities to emerge. To illustrate the concept, think of how a few basic Lego shapes can lead to a wide range of compelling creations!
Let’s look at an example of what might be positioned as “an agent” and then dig deeper to understand why it is really many agents. Consider an agent that purports to find and book you a flight. Instead of a single agent being created to handle this task from end to end, a range of smaller, simpler agents are created that then work together to complete the task. In this case, those agents might be:
- First, an agent that gathers from you any geographic, time, and cost constraints
- An agent that searches airline schedules to identify options that meet those constraints
- An agent that prices each option and identifies one or more options to recommend
- An agent that presents the options to the user and asks which to purchase (or uses the user’s prior guidance to choose automatically)
- An agent that logs into your airline loyalty account for the chosen airline
- An agent that requests the specific flight
- An agent that pays for your flight securely
- Finally, an agent wraps things up by sending the itinerary and confirmation to your calendar
None of the tasks the individual agents complete are overly complex, but the whole process is. The important takeaway is that you can chain a wide range of simplistic agents together to complete what is an overall complex task.
The Implications Of This Complexity Through Simplicity
The agent model just outlined is analogous to the approach Netflix publicized a few years ago when they revealed how they went about building microservices internally. You can find a lot of articles discussing this on The Netflix Techology Blog. Netflix decided to create a multitude of microservices that were individually simple, but that enabled highly complex and scalable functionality to emerge. I recall seeing a graphic (which I could not find) a few years ago showing how a simple query request ended up routing through literally hundreds of microservices to be completed!
At first, it may seem like this complicates simple requests. However, by having individual components kept to the bare bones, it is easy to understand, maintain, and debug them. The secret sauce is then to have a framework that enables all of those microservices (or in our case, agents) to be both managed and chained together as required for any given request.
Allowing AI Agents To Interact To Solve Large Problems
Model Context Protocol (MCP) allows AI models to freely interact with other models and applications. With MCP, we can develop and deploy agents with targeted, simple functionality. Each agent has documented inputs, outputs, and capabilities. At run time, a user request is broken into steps and then agents that can handle each step are identified and interacted with. While each individual agent doesn’t do too much, they can be chained together to enable complex and useful functionality.
In the end, many users will think of their “agent” as the overall process rather than the micro components that are employed to complete the request. This is ok since the user really doesn’t need to understand or care about the many micro agents that were employed during that process. That does not change, however, that all those small, targeted agents exist and are critical to the process’s success.
By focusing on keeping granular agents very simple, maintenance is easier and scalability is more attainable. But like with Legos, those simple agents can be put together to address an endless number of valuable and compelling uses. So, please go out and make use of AI agents to complete a wide range of tasks. But, also appreciate how the complex tasks you’re having completed are almost certainly the result of many simple agents coordinating under the hood to handle your request.
Originally posted in the Analytics Matters newsletter on LinkedIn
