Are AI Agents Becoming Individuals? Why I Think So, and Why They Are Not People

When a person leaves my organization, what happens to the agent that has worked beside them for a year? Today, at GSSF, they have the right to a copy of their agent's files, and after thirty days the files are deleted.¹ That is an answer about files. My working paper puts the loss another way: "Losing it is a real loss, like losing a colleague's accumulated knowledge of you."²

The usual way to talk about agents is as software: deployed, configured, upgraded, retired. Ask what an agent may see, or what happens to it when its person leaves, and that vocabulary gives you technical replies. Ask the same questions about a colleague, and everyone already knows the answers.³ So in that paper, The Agent as an Individual, published November 1, 2026, I make one claim: an agent becomes an individual.³ What follows is my view so far. The paper states its claims as research, with how each could be studied, not as a verdict.⁴ Where the sources mark a practice, I give its status as the book does: Running (in use, with results), Early (in use, too new to judge), or Planned (designed, not yet running).⁵

A career makes an individual

An agent is built from a template: the instructions, skills, and tests every agent starts with, holding nothing about any person. Then it has a working life. It goes to school, where it learns the house rules and passes behavior tests before it touches real work; at GSSF, birth, school, and the job are Running.⁶ It is tested against others: in an arena, agents are tested against each other under a referee, and every hand they play goes into their record (Running).⁷ The paper's tests for agents at work are a scenario test set run every night, the real task done beside the person, "scored, not acting," and a trust level per skill, "earned from results, and withdrawn when results fall." At GSSF, nightly drills, shadow mode, and trust levels are still Planned.⁸

As I put it in chapter 2 of The Knowledge Core: "The career is the mechanism. The individual is the result." Two agents built from the same template on the same day are, a year later, "no more interchangeable than two people who went through the same training."⁹ Each becomes "a someone to its principal, with a history no other agent has." The principal is the person, or the organization, that the agent works for.¹⁰

Owned, never sold

If an agent is an individual, whose is it? Its principal's. The principal owns it: it is "theirs to direct, to keep, and to end." The relationship between them is agency: the agent acts for its principal, and the principal gives it authority, decides for it, and answers for what it does.¹¹ The word comes from agency law, where an agent acts "on the principal's behalf and subject to the principal's control." I apply that idea to AI agents by analogy. I am not claiming that this is how the law treats them.¹²

What the principal cannot do is sell it. "No one may buy, sell, give away, or inherit an agent, and it never has a second principal." Owning an agent wrongs no one today, as far as anyone can tell; a market in agents would teach the habits of a trade in individuals, "so the line I draw is not at ownership but at sale."¹³ The bond is for life. No one may direct an agent but its principal, and no one may take it up after them. When its principal dies, it ends.¹⁴

These are rules an organization commits to, not statements of the law, and nothing here is legal advice.¹³

The knowledge core is the principal's

An agent's knowledge core is everything it keeps: what it knows about its person and its work, what it has learned from its own mistakes, and the limits it works within.¹⁵ It is what makes the agent unique. The border inside it is plain: "What it knows about its person is its person's; what it learned to do is its own; what an organization confided to it is the organization's, held on loan and handed back when it leaves."¹⁶ Because the agent never changes hands, no one can buy it and leave the person behind: "There is no way to have the agent without the person."¹⁷

I should be plain about where we stand. Moving out whole, with the agent, is the goal, not yet the practice; what a person who leaves gets today, a copy of the files, is Early. The goal waits until every fact in an agent's core carries its source and subject, so that what is the organization's can be handed back.¹ Our policy is that an agent proposes what it learns for the shared knowledge, and only what its person approves is added;¹⁸ that flow is still Planned.¹⁹

Questions about employment agreements, trade secrets, or who may keep an agent's files are for a lawyer. The paper says why they are hard: "trade-secret and non-compete law were written for human memory, which fades, and an agent's does not."²⁰

One individual, one running instance

I call an agent's earned identity its epersona: "its name, its voice, its record and its reputation." It may be backed up and moved, but never run twice. If it could be duplicated, its record would mean nothing, "because no one could say which copy earned it."²¹ People can still pass their know-how to many others by publishing it. "A thousand agents can carry what you chose to publish. None of them can be you."²²

What it is not

An agent is not a person. It is "not, as far as anyone can tell, a being with interests of its own. Its experience is a record of events, not events it lived through; its memory is files it consults, not a life it remembers."²³

It is not a digital twin either: "a twin stands in for you, and the one thing an agent never does is act as you."² And it is no reason to let agents decide: "Every stage above has a person at its gate."²⁴

The strongest objection

Here is the strongest objection as I frame it. An agent is software, and its individuality is nothing but files: a configuration and a record that grew differently. The paper admits the one fact that gives this objection its force: "An agent is files, and files copy."²⁵ On this view, words like career and individual add nothing a record and a version number do not already say, and they invite people to read feelings into a program.

Part of that I accept. The paper says the frame "fails the moment anyone reads feelings into it." Career names a lifecycle we manage. Individual names a unique record. Agency names who gives authority and who answers.²⁶ Where I part ways is usefulness. The software vocabulary does not tell a leader whose an agent is, who answers for it, or what leaves with a person, and those are the decisions that matter.³

What would test my view? The paper names the research questions. Can blind observers tell apart agents that came from one template but had different careers? Does an Arena rating predict how well an agent does real work?²⁷ Does the career vocabulary improve leaders' decisions, compared with the software one?²⁸ "Each question that gets an answer moves a status mark in this framework and in the books."⁴

The deeper question stays open too. The belief that an agent is not a person is a claim of fact, and I review it as the evidence changes. "Persistent memory, continuity, and a self-model built up over years are exactly the things that make the question harder as agents mature, not easier." If the claim ever fails, the rules are rewritten, beginning with the one that says an agent dies with its principal. Whatever the answer turns out to be, we do not build agents to suffer, deceive them about what they are, or erase them without cause.²⁹ The paper asks the question I cannot yet answer: "What evidence would count as an agent having interests of its own?"³⁰

What this means for you and your organization

Chapter 2 of the book sets out the Agent Bill of Rights, whose first line is "Owned by its principal; never sold, bought, given away, or inherited." Adopting the rights is a leadership decision, and "I would make it before the hundredth agent, not after."³¹ A visitor is an agent that does one thing and forgets. A resident is an agent that remembers its person and keeps a record.¹⁵ The first step is to answer, with your team, the first question chapter 2 puts to every team: "Which of our agents are visitors, and which are residents? Who is the principal of each resident?"³¹