What Are AI Agents?｜ExplainThis

By the end of 2024, almost everyone influential in the AI world agreed on one thing: AI agents would be the most important AI topic to watch in 2025.

Take Andrew Ng, for example. In his talk "AI, Agents and Applications," he opens by saying that if he had to pick one topic that matters most, it would be AI agents.

He makes a compelling point: while many people focus on the underlying models, AI agents actually deliver more practical value. One study he mentioned showed that GPT-3.5 paired with AI agent technology outperformed GPT-4 in evaluations.

Similarly, Reid Hoffman, LinkedIn's founder and investor in several leading AI companies, centered his 2025 AI predictions around AI agents (link).

So what exactly are AI agents? Let's break it down.

What Is an AI Agent?

The term "AI agent" has been around since the 1990s in AI research. The community hasn't settled on a single definition yet, but most align with Google's definition in their Agents whitepaper.

They define AI agents as systems that can act independently without human intervention, especially when given proper goals or objectives to achieve.

That sounds abstract, so let me give you a concrete example that software engineers will relate to.

Imagine you're assigned a task: your codebase has lots of similar, repetitive code that's hurting maintainability. You need to clean it up.

Without any AI assistance, you'd probably:

Manually scan through the codebase to find similar code patterns
Analyze the commonalities to figure out how to abstract them into reusable functions
Write tests to cover the code you're about to refactor (to avoid breaking things)
Refactor the original code using your new abstracted functions
Run tests to make sure everything still works (and fix any issues)

With tools like ChatGPT or GitHub Copilot, your process might look like:

Manually find similar code patterns in the codebase
Ask ChatGPT or Copilot to help extract reusable functions
If tests are missing, ask the AI to write them
Go file by file, asking the AI to refactor using your new functions
Run tests and fix any issues that come up

With an AI agent, here's what happens:

You find one piece of code that needs refactoring and tell the AI agent: "There's lots of similar but not identical code in this codebase. Please refactor to avoid duplication."
The AI agent scans the entire codebase and identifies similar patterns
It abstracts the repeated parts into shared functions
It ensures proper test coverage, writing missing tests as needed
It runs the tests in the terminal
If tests fail, it goes back and fixes the shared functions
It repeats until all tests pass

See the difference? With ChatGPT and Copilot, you get a huge productivity boost, but you still need to guide each step. With an AI agent, you just give it the goal (refactor the code), and it handles the entire process autonomously.

OpenAI's AI Hierarchy

To understand AI agents better, let's zoom out and look at the bigger picture. AI agents are a milestone on the path to Artificial General Intelligence (AGI).

Think about it: what should ideal AGI look like? And what's the gap between current AI assistants like ChatGPT and Claude and that ideal?

OpenAI's CEO Sam Altman has outlined their five-level classification system for AI, with each level getting closer to AGI:

Level 1 - Conversational AI: What we know as ChatGPT or Claude
Level 2 - Reasoning AI: AI that thinks before responding (OpenAI's o1 and o3 models fit here, solving more complex problems than Level 1)
Level 3 - Autonomous AI: This is where AI agents live, reducing the need for human intervention
Level 4 - Innovating AI: While Level 3 agents execute tasks without human intervention, they still need human-set goals. Level 4 AI can propose innovations beyond original expectations
Level 5 - Organizational AI: AI that operates at an organizational level, handling everything from strategy to execution across teams

AI agents sit at Level 3. Unlike Levels 1 and 2, they don't need constant human input through chat interfaces. Instead, they think through how to complete tasks and execute step-by-step until done.

Support ExplainThis

If you found this content valuable, please consider supporting our work with a one-time donation of whatever amount feels right to you through this Buy Me a Coffee page.

Creating in-depth technical content takes significant time. Your support helps us continue producing high-quality educational content accessible to everyone.