How to build an AI Agent
🧠 Knowledge Series #56: Your ultimate guide to AI Agents. Technical approaches explained, practical use cases and tools for product teams.
🔒The Knowledge Series is available for paid subscribers. Get full ongoing access to 50+ explainers and tutorials to grow your technical knowledge at work. New guides added every month.
Hi product people 👋,
If you believe the hype then 2025 is going to be the year of the AI agent. And for the purposes of this Knowledge Series, we’re going to believe the hype.
Right now, investors and tech leaders are predicting that the AI agent industry is likely to disrupt much of what we do every day at work, with Andreessen Horowitz recently suggesting that vertically integrated AI products could eventually become a bigger industry than SaaS, and Microsoft's CEO recently saying that in two years, building agents will be as common as creating a spreadsheet.
These predictions may or may not prove true but in this piece we’ll help you to understand what AI agents are and how product teams can build their own so that you're up to speed regardless.
We'll explain in simple terms the key technical things you need to know about how AI agents work, along with some real-world examples of AI agents in action from companies like OpenTable, Ebay, Uber and more - as well as some tools you can use to get some hands-on experience.
Coming up:
What are AI agents?
AI agent architecture explained - the essential bits worth knowing
How agents are built: 3 common technical approaches for product teams
What is Langchain and Microsoft Autogen?
Practical examples of how companies including OpenTable, Uber and more are using AI agents
Tools you can use for building AI agents
What are AI agents?
An agent is simply a program that can act on its own with some degree of autonomy to achieve the goals you set it.
The word “agent” has become so common in recent weeks that many seasoned engineers roll their eyes a little when they hear it. And part of the reason for that is because you could argue that “agents” have been around for a while - in the form of bots.
But what sets agents apart from traditional bots is that AI has given bots a new series of powers which provide them with the autonomy to make intelligent decisions of their own. A bot armed with LLMs, RAG, Machine Learning and other capabilities is far superior to a traditional bot - and agent is perhaps a more suitable word for what these things are capable of. For now at least.
At their most basic level, AI agents can perform predefined tasks under strict human supervision. These systems can follow a set of instructions and execute tasks with no decision making and are best designed to handle repetitive tasks.
Agents with conditional autonomy can perform tasks autonomously but require human intervention when faced with situations beyond their capabilities.
And agents with the highest levels of autonomy operate entirely independently, making decisions and adapting to new scenarios with no human intervention.
AI agent architecture explained
We’ll explore some of the most common approaches to how AI agents are practically built and used in the real world later, but before we dig into the details of those approaches, it’s important to understand the core technical elements of AI agent architecture and how agents are able to work with high degrees of autonomy.
Here’s a basic diagram which outlines the most important parts of AI agent architecture: