What Is an AI Agent?
If you want to understand important AI concepts, you'll want to get familiar with AI "agents." In the fast-moving world of AI, this particular term has had quite the ride. At this time last year, "agent" was mostly a buzzword: lots of promise, but not yet much substance. But by the end of 2025, agents were delivering on all the hype, and then some.
What is an AI agent? A good place to start is the familiar meaning of "agent": someone who acts on behalf of someone else. A traditional chatbot like ChatGPT is not an agent because it can't do anything other than write text in response to your prompts. But if you give an LLM access to a set of tools and instructions about how to use them, you can give the model the capacity to act—to do something on your behalf.
The most widely used type of AI agent is certainly the coding agent. Last March I wrote about my discovering how much more powerful a coding agent is than a chatbot. These tools (Claude Code from Anthropic and Codex from OpenAI are the best-known examples) have completely transformed software development. LLMs can write computer code at least as well as they can write prose. So if you give them the tools to write code on your behalf, you transform your part in the software development process. When you work with a coding agent, programming is no longer about writing code yourself; it's about giving good instructions to the coding agent about what you want the software to do.
It's hard to overstate how powerful these tools are.
And coding agents are just the beginning. We're now seeing an expansion of agents into tasks beyond coding. Anthropic recently released a product they call Claude Cowork, which is essentially a version of Claude Code that's intended to be more accessible to non-programmers for use in non-programming tasks. Last week, a project called OpenClaw exploded onto the AI-techie scene promising to be an all-encompassing AI assistant. The project homepage prominently claims that it "Clears your inbox, sends emails, manages your calendar, checks you in for flights." Tools like these are gaining traction quickly.
I'll finish this introduction to the concept of agents by pointing out the dark side. Having powerful tools act on your behalf is great when they do what you want them to do, but what about when they don't? With agents, things can go wrong not only through old-fashioned bugs in the code, but also via whole new classes of failures such as prompt injection that are unique to agents. It's a dangerous dynamic: agents are most powerful when they operate with the greatest freedom. (I experimented with OpenClaw but set it aside largely because I felt it was too cavalier with the capabilities it gave to the agent.) It's likely that the road ahead will include some painful lessons.
(PSA: If you experiment with agentic tools like the ones I've linked above, tread cautiously. Don't give the agent access to any more data than what it needs to accomplish its task. Read the vendors' security warnings carefully.)