What an AI agent actually is

I’ve built several generative AI tools and a few agents. Still, there’s one question that keeps coming up: what actually is an agent? From a user’s point of view, the concept is usually intuitive: a conversational agent that replaces human functions like sales or support. But from a technical standpoint, the definition goes deeper and is relatively new. For a long time there was no formal description, which has caused confusion. Lots of developers call a simple script that uses the OpenAI API an “agent”, when it really isn’t one.

My first attempts at conversational agents

My first attempt at building a conversational agent was about 10 years ago, when Microsoft launched Cognitive Services on Azure. Back then, I built a chatbot hoping to make it interactive and intelligent. The result was a total failure. The technology was at a very early stage and the models couldn’t reason or keep context like today’s do. Still, that experience sparked my curiosity about how conversational systems could evolve into what we now call agents.

Types of agents

Although conversational agents are the best known, they’re not the only ones. There are agents for data analysis, search, code automation, planning or monitoring. Still, conversational agents are a good starting point for understanding the concept, since they let us see how a model interacts, reasons and acts with some degree of autonomy.

Core characteristics

For a system to be considered an agent, it has to meet certain characteristics. Today they’re built on LLMs (Large Language Models), although other technologies could be used in the future. First, an agent makes calls in a loop to external or internal APIs. This cycle of reasoning and action is what makes up its behavior. What we perceive as a conversation or problem-solving is really a sequence of chained calls.

Second, an agent uses tools (tools). These tools give it specific capabilities, like searching the web or doing calculations. LLMs aren’t good at everything and don’t have up-to-date context, so they delegate certain tasks to external components. For example, even though a model understands how arithmetic works, it isn’t especially accurate at calculating. Just like a human uses a calculator to get past their limitations, the agent turns to tools to perform better.

Another essential piece is memory. Memory lets it keep the context of the conversation or previous actions, enabling continuity and coherence across the following iterations. In short, an agent:

• Makes calls in a loop until it tries to solve a problem.
• Uses tools to extend its capabilities.
• Uses memory to remember and reason progressively.

Workflows, algorithms and autonomy

This is where the concepts get mixed up. Many systems presented as agents are actually workflows or algorithms. The difference is subtle but key. A workflow is a sequence of predefined, controlled steps, where each action depends on the previous result. There are tools designed for building these kinds of flows, like n8n, Zapier or the new OpenAI Agent Builder, which let you chain tasks and orchestrate APIs in a controlled way. These systems are useful, but they’re not agents.

Anthropic’s researchers explain it clearly: an agent is not a workflow. In a workflow, the number of steps and the decisions are predetermined. In an agent, the model decides how many times to act, which tools to use and when to stop. It has autonomy. That autonomy is what makes it more than a sequence of calls. Instead of running A → B → C, the agent can decide to do A → B → A → C → D until it finds a solution. The programmer is no longer in control; the model is, within a defined framework.

Why everyone says they’re building agents

Today, the term “agent” has become attractive. It sounds innovative and complex, so lots of people and companies use it even when it doesn’t apply. In many cases this comes from not knowing better: people think that calling the OpenAI API or building a flow with several calls already turns a system into an agent. But there’s also an aspirational or marketing component: “we’re building agents” sounds better than “we call the ChatGPT API”. Even among colleagues or bosses, using that word gives a sense of modernity or sophistication that doesn’t always reflect the technical reality.

Building with AI vs building AI

Another common confusion is believing that building agents means being an expert in machine learning. They’re different things. Building with AI means using already-trained models —like GPT or Claude— to create intelligent solutions. Building AI, on the other hand, means designing and training those models from scratch, working with data, architectures and deep learning techniques. Most people building agents are in the first group: they’re developers building intelligent systems on top of existing models, not scientists training new ones.

Conclusion

An agent is not a chatbot or an algorithm. It’s a system that reasons, remembers, decides and acts within a defined framework, using tools to make up for its limitations and APIs to take actions. What sets it apart isn’t just its ability to hold a conversation, but its autonomy to decide how to solve a problem. And even though the term is trendy, understanding these differences is key to building real solutions, not just systems that “look like” agents.

Thanks, collaboration and next steps

Thanks for reading this far. I enjoy this topic because it’s been a personal journey of many years, from my first failed experiments with chatbots to understanding how today’s agents work. I’m writing a new post to dig deeper into an agent’s working framework, or universe, and also to reflect on how to build these kinds of projects solo or as a team. As always, my passion for teamwork gets me thinking about how these new technologies can inspire new ways of collaborating and learning together., where I’ll share practical lessons and thoughts on how to define an agent’s boundaries, capabilities and autonomy based on my experience and the content I’ve consumed.

Some references:

• https://www.youtube.com/watch?v=LP5OCa20Zpg
• https://blog.langchain.com/what-is-an-agent/