Jargon:AI agent
2 min read
Say this at the party
“An AI that does things, not just says things: you give it a goal and it uses tools, checks its work and keeps going until the job is done.”
In plain English
A chatbot answers. An agent acts. You give it a goal, like "book a table for Friday", and it works out the steps itself. It can use tools such as a web browser, email or code.
After each step it looks at what happened and decides what to do next. Anthropic, an AI company, puts it simply: agents are usually just a language model (the brain inside chatbots like ChatGPT or Claude) using tools in a loop.
The word is slippery. Anthropic says it "can be defined in several ways", from fully independent systems to fixed step-by-step routines. So when someone says "agent", ask what it actually does on its own.
Where it came from
Researchers were writing about agents long before chatbots. In 1995 two UK academics, Michael Wooldridge and Nicholas Jennings, said agents had already "become important" in AI. They listed traits like acting without a human stepping in and taking the first step towards a goal. The chatbot meaning came later. In December 2024 Anthropic's guide "Building effective AI agents" gave the industry a widely used working definition.
See it in action
In January 2025 OpenAI launched Operator, which took over a web browser to reserve restaurants and shop online, clicking buttons and filling in forms like a person would. It asked you before placing an order, and handed back to you when it got stuck on a "prove you're not a robot" puzzle.
What people get wrong
- Not everything called an agent is one. Gartner calls this "agent washing": old chatbots with a shiny new label. It reckons only about 130 of the thousands of "agentic AI" sellers are the real thing.
- Agents don't have minds or goals of their own. They're a language model plus tools in a loop, and the goal comes from you.
- More independence isn't always better. One wrong step can feed the next, so mistakes pile up. Anthropic advises the simplest setup that works, which is often no agent at all.
- They're not ready to replace staff. In a 2024 Carnegie Mellon test set in a pretend company, the best agent finished only 24% of tasks on its own. Even fan Andrej Karpathy says to "watch them like a hawk".
Go deeper
- GuideAnthropic, 2024Building effective AI agentsThe working definition most people now use, plus why simpler is often better.
- PaperWooldridge and Jennings, 1995Intelligent agents: theory and practiceProof the idea is decades old, from two UK researchers.
- ArticleMaxwell Zeff, TechCrunch, 2025OpenAI launches Operator, an AI agent that performs tasks autonomouslyA browser agent booking and shopping, and where it got stuck.
- PaperXu et al., 2024TheAgentCompany: Benchmarking LLM Agents on Consequential Real World TasksAgents tested on pretend office work. A useful reality check.
- ArticleGartner, 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027Where "agent washing" comes from, and why many projects stall.
- ArticleMatthias Bastian, The Decoder, 2026Former Tesla AI chief Andrej Karpathy now codes "mostly in English" just three months after calling AI agents uselessA leading researcher's quick change of heart about coding agents.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.