Principal

What it means

An agent's principal is the person it works for. In The Knowledge Core, the principal is "the person who gives it its work and answers for what it does."¹ The book coined this use of the word for AI agents, and it builds on centuries of law and fifty years of economics.

Here is an everyday example. A sales manager tells an AI agent to draft renewal offers for her ten largest customers. She is the principal. The agent takes its work from her, and if it offers a discount nobody approved, she is the one who answers for it. The agent doesn't answer for it, and neither does the software company that built it. The idea behind the book's rule that a person stays in the middle of every chain of command is simple: every agent should have a principal like her, and that principal should be a person.

Where the word comes from

The word comes from law. In agency law, the body of law that governs one party acting for another, a principal is someone "that authorizes another person or entity (known as an agent)," who then acts "on the principal's behalf and subject to the principal's control."²

In the 1970s the word moved into economics and political science. Barry Mitnick, one of the scholars involved, writes that the people who first developed a theory of agency "were Stephen Ross and Barry Mitnick, independently and roughly concurrently." Ross's paper was published in 1973. Mitnick also notes that "the law of agency had a role in the naming of the theory of agency."³ The question at the center of that field is how a principal gets an agent to act in the principal's interest. That question is one reason the word suits AI.

AI companies have now taken up the word. In January 2026, Anthropic released a revised version of the document that guides its Claude models, which it calls a constitution.⁴ That document uses "principals" for the people Claude works for.⁵ The Knowledge Core uses the word in this sense too, and it adds one requirement: the principal must be a person who answers for the agent's actions.¹

How others define it

• Agency law (Cornell Law School's Legal Information Institute): "In agency law, a principal is a person or entity that authorizes another person or entity (known as an agent), to act on the principal's behalf and subject to the principal's control."²
• Anthropic: "We use the term “principals” to refer to those whose instructions Claude should give weight to and who it should act on behalf of." The document also says: "At the moment, Claude’s three types of principals are Anthropic, operators, and users."⁵
• **The Knowledge Core:** "the person who gives it its work and answers for what it does."¹

Where they agree and differ

All three agree on the core idea: a principal is the one the agent acts for, and the agent should follow the principal's direction. Anthropic separates principals from people whose interests still count. It describes "those whose interests Claude should give weight to, such as third parties in the conversation" as a different group.⁵ An agent should treat outsiders fairly, but it takes instructions only from its principals.

They disagree about who can be a principal. In the law, a principal can be a "person or entity," which includes a company.² Anthropic's list includes Anthropic itself and the businesses that build on its platform.⁵ The Knowledge Core asks for something narrower: a person.¹

The strongest objection is a fair one. Inside a large organization, the agent really serves the company, and naming one person as its principal could be a fiction or put unfair blame on that person. The law accepts companies as principals for good reasons. The book's answer is that a company can't be asked a question on a Tuesday afternoon, and a person can. Naming a person doesn't remove the company's responsibility. It makes sure someone can actually be asked about the agent's work.

In the news

In August 2026, the Ninth Circuit Court of Appeals overturned an order that had stopped Perplexity's AI shopping agent, built into its Comet browser, from shopping on Amazon. Engadget reported that "the appeals court determined it was the Comet user who was accessing Amazon's" computers.⁶ Put simply, the court treated the agent as acting for the person using it. We are fairly sure of this summary, but we are reporting the reasoning from news coverage because we could not read the opinion's text ourselves.

The best-known earlier case went the other way for a company. In 2024, Air Canada argued that its website chatbot was "a separate legal entity that is responsible for its own actions." The tribunal disagreed. It said the chatbot "is still just a part of Air Canada's website" and that "It should be obvious to Air Canada that it is responsible for all the information on its website."⁷ The airline had tried to say the agent answered for itself, and the tribunal held that the airline did.

What it means for your organization

These two cases point the same way. When an agent acts, courts look for whoever it acted for, and that party answers for what it did. If you don't name a principal for each agent, a court or a customer will decide who it was, and it will usually be your organization. Before you deploy an agent, write down whose instructions it follows, whose interests it must respect, and which person answers for its work.

One question to ask your team: For each AI agent we run today, can we name the one person who answers for what it does?