AI agent

What it means

An AI agent is software that gets a goal, works out the steps on its own, and takes action to reach it. A chatbot answers your question and waits for the next one. An agent keeps going: it can search, fill in a form, send a message, or update a record until the job is done.

Here is an everyday example. You ask an assistant to "find me a flight to Chicago on Tuesday under $400 and hold it." A chatbot gives you a list. An agent checks the airlines, compares the fares, picks one, and places the hold. In between, it makes small decisions that you never see.

The companies that build agents put it the same way. Anthropic describes agents as "systems where LLMs dynamically direct their own processes and tool usage."¹ (An LLM, or large language model, is the kind of AI behind tools like ChatGPT and Claude.) IBM says an agent "autonomously performs tasks by designing workflows with available tools."²

Where the word comes from

Computer scientists used "agent" long before today's AI. In 1995, Michael Wooldridge and Nicholas Jennings opened a widely cited paper with this line: "The concept of an agent has become important in both artificial intelligence (AI) and mainstream computer science."³ We could not confirm who first used the word in this sense, so we do not name anyone.

The meaning changed when large language models arrived. An agent used to be any program that senses its surroundings and acts on them. Today it usually means a language model with tools added around it. In January 2026, the U.S. National Institute of Standards and Technology (NIST) described AI agent systems as made of "at least one generative AI model and scaffolding software" that gives the model tools.⁴ The same notice says "Other terms used to refer to AI agent systems include AI agents and agentic AI."⁴

How others define it

• Anthropic: "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage."¹
• IBM: "An artificial intelligence (AI) agent is a system that autonomously performs tasks by designing workflows with available tools."²
• NIST (U.S. government): "AI agent systems are capable of planning and taking autonomous actions that impact real-world systems or environments."⁴
• Our view: The Knowledge Core puts it this way: "An AI agent works for one person, its principal, who gives it its work, decides for it, and answers for what it does."⁵

Where they agree and differ

All of them agree on the core idea: an agent chooses its own next step and acts, instead of following a fixed script. Each one stresses something different. Anthropic stresses the model's control over the process. IBM stresses tools. NIST stresses effects in the real world and adds a warning: "They can be deployed with little to no human oversight."⁴

Our definition is the only one built around a person. The others describe what the software does. Ours describes who it works for and who answers for it. That is a choice about leadership, not a technical claim.

The strongest view against ours deserves a fair hearing. It holds that agents should be treated as risks to contain, not as workers to lead. Microsoft's security chief told TechCrunch, "Unfortunately, frontier models, regardless of their origin, cannot be trusted."⁶ We take that seriously. In our view, it supports a named principal rather than replacing one: someone still has to decide what an untrusted agent may do. Readers may reasonably disagree.

In the news

On October 27, IBM announced tools that give every agent a verified identity. The record shows the agent's owner, its status, and the authority it has been given, along with a "kill switch" that can shut any agent down.⁷ IBM's release says plainly, "Enterprises are deploying agents faster than they can verify them."⁷ The release also cites a Gartner projection that by 2028 the average Fortune 500 company will use more than 150,000 agents.⁷ That number is a forecast relayed by a vendor, so treat it as a sign of the direction, not as a measurement.

A week earlier, TechCrunch reported that Microsoft now treats agents as a possible insider threat and has launched Guardian Agents, "a new feature intended to monitor the AI agents a company uses."⁶ TechCrunch also noted that Microsoft sells these security products.⁶ We are confident in both reports as accounts of what the companies said. Neither one shows how well the tools work.

The two stories point the same way. The largest software makers are now building ways to record who owns each agent and to stop one that goes wrong.

What it means for your organization

An agent's choices are its own, so the important question is not only what it can do. It is also who answers for what it does. The Knowledge Core argues that each agent should have one principal and that a person should stay in the middle of every chain of command.⁵ IBM's "owner" field and Microsoft's monitoring are tools for that idea. They do not replace it.

One question to ask your team: For each AI agent we run today, who is the one person who answers for what it does, and could they stop it this afternoon?