AI Agents Simplified

The article explains how AI agents operate using concrete examples and the terminology related to them.

One obstacle to leveraging generative AI and especially its latest development AI agents in organizations is a lack of understanding of how they work. It's difficult to push new initiatives if there's no shared understanding of what you're doing and why.

A lack of shared understanding and vocabulary creates uncertainty that can stifle the entire conversation. If, as a leader, you don't understand what's being discussed, you'll probably prefer to drive forward other topics.

But don't worry! When we stay away of technology jargon, AI agents are an easy concept to understand. In this article, I'll walk through the topic with concrete examples while also explaining terminology from the world of generative AI and AI agents.

Stages of Leveraging Generative AI in Organizations – Are We Already Using AI Agents?

Before diving into the main topic, let me first clarify what I mean by AI agents in this context. I've divided the stages of leveraging generative AI as follows: Shadow AI, generic and customized chat-based assistants, and finally simple and collaborative AI agents.

Stages of leveraging Generative AI.

In this blog, I primarily address the last two categories, but in practice, most chat-based assistants today – including the examples mentioned above – actually consist of multiple agents working behind the scenes. When we give them a prompt, they formulate a plan on how to respond and utilize various tools (such as web search) to generate the desired outcome. Both of these are characteristic features of AI agents.

In other words, most of us are already knowingly or unknowingly using AI agents, but in this article, we'll focus on their use as separate, autonomous implementations.

Agent Is "Just" a Loop

Generative AI is based on so-called large language models (such as Anthropic's Claude and OpenAI's GPT models), which have been trained on enormous amounts of data to interpret the prompts they receive and reason out answers to questions or tasks given to it.

And in simplified terms, an AI agent is "just" an iterative loop of such prompts and reasonings. However, one possible outcome in this loop is the decision to use a tool that has been made available to the agent. This is one of their distinguishing features. Another characteristic is the instructions given to the agent, which guide its behavior.

When we give an agent a prompt, or a task, the following reasoning chain occurs:

1. The agent uses large language model to interpret the given prompt and instructions, and based on them, formulates a plan on how to complete the assigned task, taking into account the available tools

2. If the conclusion is that it needs to use a tool, it calls that tool, and upon receiving a response, it reasons again about how to proceed with the plan

3. The agent continues this "plan - tool - reasoning" loop until it considers the task at hand to be done

4. After which it returns a response according to its instructions. For example, describing what it has done or providing an answer to the question provided in the prompt.

The image below illustrates this loop.

AI Agent reasoning chain.

Wait a moment... that doesn't quite make sense yet, and what exactly are these tools?

Well, instead of trying to answer right away, let's take a thought exercise here – what would a car wash look like as an AI agent?

What Would a Car Wash Look Like as an AI Agent?

A car wash gives us an excellent way to think about AI agents through a familiar concept. If it were an AI agent, its instructions could be to wash the car according to the customer's wishes as efficiently as possible, along with some specific work instructions like that it should dry the car before waxing. Its tools could include manual and machine pre-wash, brushless and brush wash, drying, machine waxing, vacuuming, and interior cleaning.

Now imagine that the prompt given to this agent is a car covered in mud and salt after a spring trip to the cottage. The customer's wish is to get the exterior clean quickly.

Based on the prompt and instructions, the agent would create a plan: machine pre-wash, brush wash, drying, waxing, and another drying at the end. Then it would send the car to the machine pre-wash, and upon receiving a response that the biggest dirt has now been washed off, it would continue implementing the plan to the brush wash, and so on, until finally returning a clean car to the customer.

What Would a Car Wash Look Like as an AI Agent?

Aha... it's starting to make sense! But if we now think about the actual use environment for agents, what would these tools, instructions, and prompts be in that context?

AI Agents in Real Life

That's a tricky question because they can be almost anything. A tool can be as simple as a "calculator" or getting the current date and time for the agent to use.

Although large language models contain an incredible amount of information, they're "dumb" in some simple matters. They don't know, for example, what day it is. Now consider an agent that processes purchase orders. It needs to provide a delivery date, which requires knowing the current date as a starting point. The answer is a simple tool that the agent can query

On the other hand, tools can be quite complex, such as a web browser that allows the agent to perform actions on the internet like a human would. Or it can be a tool that makes an HTTP request to fetch or push information to any publicly available application interface on the internet.

In summary, a tool can be anything you can implement as code. When discussing tools, you might also encounter the abbreviation MCP (Model Context Protocol), which is an open standard for agents to discover and use tools. With MCP, different parties can "publish" tools related to their own services or information systems, and instead of separately telling the agent which individual tools are available, it can discover and use them on its own through MCP... apologies, I went a bit too deep here. Let's get back to a practical example.

Let's take the above mentioned purchase order processing as an example. An agent designed for this purpose would have instructions to compare incoming purchase orders against the company's product information and determine which products (SKUs) the order concerns, check inventory levels, estimate delivery dates, calculate freight costs, and process it and confirm it to the customer with a shipment tracking code.

Its tools would include the necessary interfaces to CRM and ERP systems, logistics partner systems, email, and so on. And just like in the car wash example, when it receives an order as a prompt, it formulates a plan to process it and starts to execute.

At this point, a reasonable critic would ask Do we really need an agent for this? Couldn't this be handled with just traditional integrations?

To Use an Agent or Not to Use an Agent, That Is the Question

And that's a valid question. If the environment and incoming tasks are highly standardized, the benefits of an AI agent remain minimal, and it's worth looking at other alternatives. But if the incoming tasks are not standardized (if every customer orders using a different form, with their own descriptions and product codes, and provides all kinds of instructions) then the benefits of agents become apparent as they can interpret purchase orders using large language models and adapt to the task at hands.

Or if your processes have many variables, the rigidity of deterministic methods (such as robotic process automation, RPA) can become an obstacle. Or what if all the necessary information isn't in the order? No worries, the agent can ask missing information from the customer via email and continue processing the order once it receives a response.

AI agents are suited for situations that require multi-step reasoning, adaptation, and/or frequent interaction with humans (or in the future, with other agents). Related to this, I also recommend reading Aines COO Markus Mertanen's article on How to Identify the Best Workflows for AI.

A Brief Word on Multi-Agent Systems

What are multi-agent systems? In all simplicity, they are systems consisting of multiple individual agents, each with their own specific role and tools, that can communicate with each other.

Most commonly, this approach involves one supervisor agent with several specialized sub-agents beneath it. One practical reason to use multiple agents is that the more complex the workflow, the greater the probability that the large language models behind agents will start making errors. In other words, by dividing tasks among multiple agents, we can apply a divide-and-conquer principle to implement more complex workflows with their help.

I hope this helped you understand better the world of AI agents, and if you'd like a 15–30 minute briefing on the topic for your team or executive group via Teams, drop me a message and we'll schedule a time.

-Jarkko