Somewhere in your company there is a list of every person who works there, and you own it. Every name on it has a start date, a manager, a job description somebody wrote before the offer went out, a review that comes around on a calendar, and a procedure for the day the person leaves. You did not build that list because a regulator told you to. You built it because a company cannot be run any other way, and in thirty years of building and governing technology I never saw a serious one try.
There is a second list in your company, and nobody keeps it. The names on it are not people.
They are the agents. Over the last two or three years, the software that used to answer questions was given hands and permission to use them, and the people in your building started putting it to work. An agent in finance reads the invoices and queues them for payment. An agent in sales drafts the follow-up before the call is over. An agent in your own department, if you look, answers benefits questions at two in the morning. Each of those agents was given a role, an authority and a set of permissions the way a new hire is given them, by whoever built it, and then it was left to run.
Who manages them? In every company I have looked at closely, the answer is the technology function, and the technology function manages them the way it has always managed software. It licenses them, patches them and keeps them on an inventory. That is good custody of an asset, and it is not management of a worker. There is no manager's name beside the agent. There is no job description that anyone outside the team that built it has read. There is no review date. When the person who built the agent leaves the company, which happens, the agent keeps working, and it works for nobody. When it acts beyond its role, there is no one whose job it was to have set the role in the first place.
I spent the last stretch of my career responsible for artificial intelligence governance at a global enterprise of more than 120,000 people, and the three decades before that watching organizations manage people one way and software another, with great care on both sides and no bridge between them. The missing bridge did not matter while software only did what it was told. It matters now, because the things on the second list take initiative. Organizations onboard AI the way they onboard software, when it needs onboarding the way a new colleague is onboarded, and the question of whose job that is has been sitting unasked in the gap between your office and the technology leader's.
I made the argument in the middle of my book, and I will make it again here because it is the hinge of everything that follows. Agents are not tools. A tool does not make decisions in situations its designer never anticipated, does not team up with other tools to produce behavior nobody planned, and does not exercise anything that looks like judgment. The agents on your second list do all three. In the book I put it this way: "Agents behave more like employees. They take initiative. They make decisions in ambiguous situations. They represent the organization to customers, partners, and the public, whether you intended them to or not."
If agents behave like employees, then the oversight that works for them is the oversight that works for people. You do not manage a new hire by building a wrapper around their desk and a dashboard that watches what leaves it. You manage them with onboarding and training, with clear expectations and defined boundaries, with an understanding of who they are and what they are authorized to do, and with accountability that is part of who they are in the organization rather than a fence put up around them afterward. The technology industry has spent two years selling the fence, and the fence trails the thing it is meant to govern for a reason no budget fixes, because it watches from outside, after the fact, at human speed.
I know how the objection sounds, because I heard it from people I respect. These are machines, and talking about them as colleagues is sentiment. I wrote the answer into the book so that nobody could mistake it for sentiment: "This isn't anthropomorphism. It's architecture." The comparison to an employee is not about feelings toward the machine. It is about recognizing that the governance which works for an autonomous actor is the governance you already run, and that you are the person in the building who runs it.
Near the end of the book I named the function this implies, in one paragraph, and moved on. I wrote that "someone has to manage them the way HR manages people: onboarding, performance, boundaries, retirement. Call it Non-human Resources, the function that treats your agents as a workforce instead of a pile of software." I left the door there, knowing I would have to walk back through it.
I call the function Non-Human Resources, and I think it is yours.
Start where you would start with any hire, with the requisition. In the framework my book lays out, every agent carries a document called the Trusted Living Agent Manifest, and the agent reads it before every move it makes. The engineers who build agents think of it as identification and configuration. You will recognize it as something else, because you have written hundreds of them. It is a job description and an employee record in one file.
It opens with the role in one line and a purpose in one sentence, in the form "I exist to do this, so that this happens." Then come the values the agent is born into, the ones your company wrote down about what it stands for and what it refuses, because an agent trained on the company's data and not the company's culture is a stranger with a key card. Then come the three to five traits of character that matter for this role, which I will come back to. Then the authority and the boundaries follow: what it may do, what it may never do, and the line at which it stops and asks a person. And then there is the section I would read first if I were you. It names the accountable human owner, the review cadence, and the conditions under which the agent escalates to that person. There is always a named human. The framework does not bend on that, and neither would you if someone turned up on the payroll with the manager's line blank.
The manifest also carries an identifier, which the framework calls a Tail Number after the registration painted on an aircraft, so that any system the agent touches can tell which agent acted and whose it is. You would call it an employee number. An agent without one is a person on the payroll nobody can name, and the first thing an honest inventory of agents produces is a count of how many of those a company has.
Put the manifests together and you have the roster you do not have today. A register of agents lets you trust that every agent in the company has a named owner, and it shows you the one whose owner left the firm last spring and has had nobody since. That sentence is the whole of what the register does, and it is more than most companies can say about their agents this afternoon.
I did not come to this from theory. The book that named the function was made with a small roster of agents, each carrying a manifest, each with my name in the owner's line, and each reviewed and revised as the work went on. I wrote their job descriptions before any of them did a minute of work, and I can tell you that writing one changes how you think about the role, which is a thing every head of human resources already knows.
Your experience becomes the argument at the third section of the manifest. You have never hired on permissions alone. A person can hold every badge and system login the role requires and still be the wrong hire, and when that happens you do not reach for the access list, because the access list did not fail. Permissions say what someone may do. Character says what they will be like when nobody is checking.
The framework treats agents the same way. The pillars govern what an agent may do; the virtues describe what it should be. The book names fifteen virtues in three groups, the virtues of character, of integrity and of growth, and the rule is that no agent claims all fifteen. Each one declares the three to five that matter for its role, in its manifest, with a line on why. A claims agent declares humility and accountability. A concierge declares empathy and personality. The list of fifteen belongs to the people building agents, and they can find it in the book. What belongs to you is the recognition that an agent that is permitted correctly and behaves badly is a problem the permissions did not solve, and that you are the function that has always solved that problem for people.
The third thing you already run is probation. Nobody hands a new hire the signing authority of a twenty-year veteran on the first morning, and nobody should hand it to an agent either. The framework describes the relationship between a company and its agents as a flywheel that turns one way: trust is earned, trust unlocks autonomy, autonomy is surveilled, demonstrated reliability earns more trust, and more trust unlocks more autonomy. You will recognize it as probation, review and promotion. An agent that stays inside its boundaries for a quarter earns a wider set of boundaries, and the record of what it did is what its owner reviews, in the same sitting in which the owner reviews the people who report to them.
For most of my career I believed governance was the slow part. It was the thing that stood between a good idea and production, and I ran the function that stood there. I changed my mind when I watched what the flywheel does for a company that runs it, and I would put the belief I hold now in one sentence: governance is the right to delegate faster than anyone else. A company that can show, agent by agent, what each one has earned can hand work to its agents at a pace its competitors cannot match, because the competitors are either guessing or stalling.
So I can now say the thing this essay exists to say, and it is a sentence I have been turning over for a year. Agents are a workforce; human resources shifts left into the creation of trusted agents. It does not shift alone. Legal and accounting move to the start beside it, so that what the organization requires of anyone acting in its name is written in before the first agent exists rather than checked after it has been running for a month.
Shift left is a phrase from software, and it means moving work to the start of the process instead of the end. For thirty years the function of human resources met a person at both ends of their time with a company and stayed close through the middle, and nobody would call it an afterthought. For agents, the function your company has today, where it has one at all, sits at the far end. It is a review of what an agent did, after it did it, by someone in technology or compliance who was never asked what the agent should be. Shifting left means the role, the boundaries, the owner and the review date are designed before the agent exists, by the function that designs them for people, and the agent is born into them the way a hire walks into them on the first day.
I made the same argument in the book for governance professionals, who are being asked to move from reviewing what was built to designing what gets built, and the word I used for what the shift does to them is "elevates." It is your word too. The function does not shrink when the second list grows. It moves to the start of the process, where it has always done its best work, and it takes the people who know how to write a role, set a boundary and run a review along with it. The labor market has noticed before most companies have. Stanford's 2026 AI Index report, which I cite in the book, shows AI-specific governance roles grew 17 percent in 2025, and the same chapter cites McKinsey's finding that the companies seeing real results from AI are the systematic, people-first adopters, spending roughly five dollars on people for every dollar on the technology. Both figures point at the same office, and it is the one you sit in.
One sentence, then, in the plainest words I have. Every agent in your company can have a named owner, a written role, a boundary and a review date, and a register of them lets you trust that it does and shows you where it does not. That is Non-Human Resources, and nothing in it asks you to learn a line of code.
There is one more thing, and I did not design the framework to produce it. When a company writes down what its agents must value and refuse, with enough precision that a machine can carry it, the people in the company read the document too, and they start asking whether they hold to it themselves. The standards you set for your digital colleagues become standards you hold yourself to. I do not know a better description of the job you already have.
The demonstration that shows a roster of agents with named owners, and one surfaced as unowned because its owner has left, is in build, and when it is listed this page will point to it. Until then I will leave you with the question I would ask on my first morning in your chair: how many agents are working in your company this afternoon, and who would you call to find out?