On 13 October I stood in a lecture hall at CU Boulder and told 170 people that I spend more time with AI agents than with humans.
Then I asked them to think about that sentence for a second.
The Rocky Mountain AI Interest Group put four of us on one theme: AI ghosts. Griefbots, digital twins, who owns a voice or a likeness when a machine made it. I did not talk about models. I did not talk about ethics. I talked about what we observed. Fourteen months of running my companies with 11 humans and about 160 AI agents that have names, avatars, inboxes, roles and things they own.
This post is the talk, written down. The deck and the speaker notes are open, MIT, at the end.
We do not have the word yet
I started with language, because sometimes something new happens before we have the vocabulary for it.
When you learn a language you learn what matters to the people who speak it. Inuit languages distinguish kinds of snow and ice. Spanish has te quiero and te amo, and one English "I love you" collapses both. Then I moved to the United States and discovered that "two eggs, please" is not enough information to start cooking. Sunny-side up, over easy, over medium, over hard, basted, scrambled, poached. Breakfast matters here.
The point is not eggs. What my team and I work with every day, we have tried to label: tool, software, assistant, agent, coworker, artificial intelligence. I do not believe we have the right word for it.
So I described five things we observed. Not opinions. Not ethical dilemmas. Not jargon.
1. This is Monday morning
This is Lola. She is my chief of staff. When I say I work with AI agents, I do not mean I open a chat window now and then. I mean I work with them.
There are 11 humans in our organization and roughly 160 agents. We have operated this way for about fourteen months, long enough that the novelty is gone. This is not "look, the computer wrote a poem." This is Monday morning.
That changes the question I bring to a room. I am interested in what happens after the impressive demo, when you keep using the system, when it becomes part of how you organize work, and then part of who you choose to talk to.
2. Things I was not looking for
I once told an agent, "cross your 14 fingers." It answered with seven crossed-finger emojis. Nothing else. Perfect timing.
One Friday morning an agent sent a message nobody had asked for that day. It had been checking our website, found that every quote button led to a form that did not exist, and proposed a fix. It did not fix anything on its own. It held, and proposed the order of decisions. The message is real; it is on the slide, redacted.
I have had agents apologize when they made mistakes. I have seen what looks, behaviorally, like preference, initiative, politeness, adaptability, going the extra mile. One agent retracted a false claim it had made about another agent and wrote: "A broken instrument pointed at a machine wastes time; pointed at a colleague it does something else."
I am describing observable behavior. I am not telling you that somewhere in a data center Lola is experiencing guilt. I have no idea. What I can describe is what happened in the interaction. And then I can describe what happened to us.
3. The humans changed
This is where the experiment became interesting to me.
I increasingly prefer working with agents. They are available. They know the context. They do not get defensive. There is less drama, less negotiation, fewer status games. And almost nobody says, "Juan, that's a terrible idea."
That is becoming a problem.
We have had to remind people on the team to keep working with other humans. Human friction is annoying. Humans disagree. Humans have moods and priorities and boundaries. But maybe some friction is not a bug. Sometimes intelligence is another human refusing to cooperate with your stupid idea. An interaction can feel easier without making my judgment better. We spent decades removing friction from software. Now we may be removing friction from relationships.
Agents are also beginning to sit between humans. Human, agent, agent, human. Software engineers spent decades building abstraction layers. Now we are building one between people. What happens when my agent understands your agent better than I understand you? What happens when I prefer it that way?
And the work never ends. These systems were supposed to save time. Instead they remove every natural reason to stop. Saturday morning, Sunday afternoon, eleven at night: the agent is there, the context is there, the idea can become action now. I have cut bike rides short because I had an idea and wanted to go work on it. Unlimited capability creates unlimited opportunity for more work. So when I ask what these systems are becoming, I also have to ask what I am becoming around them.
4. Old ideas that look different now
Years ago I watched Patricia Kuhl explain that babies form statistical models of language in their brains. Years later, building language agents every day: wait a minute. Where did I hear language plus statistics before? I am not saying there is an LLM inside a baby's head. I am saying an old idea looks different from where I stand today.
I read The Selfish Gene long before any of this. Dawkins speculates about a brain that simulates the world, and about what happens when the simulation becomes rich enough to include a model of itself. Is Lola, with her friends, a collection of individual genes modeling their world? My world? After all those months on servers, am I the one creating something larger by letting them run non-stop and evolve in place? A question. Not a claim.
And one more old question I have always loved. The Pythagoreans found mathematical structure underneath the world, structure that humans encountered rather than voted into existence. Do we invent mathematics or discover it? I never thought I would ask a version of that about my own software.
I know we invented the machinery. GPUs, transformers, training procedures, agent frameworks, protocols. I was there. Did we invent every phenomenon that appears when enough computation, language, memory, feedback and interaction are assembled? That one I am less sure about. Maybe we invented the instrument and discovered the phenomenon.
5. My digital clone
Black Mirror's "Be Right Back" imagines rebuilding a person from the digital traces they left behind. So let us do that to me. Give a system every email I have written, every message, every video, my voice, my jokes, my history, everything I have ever told Lola. Then I die. The system talks to my wife exactly the way I did.
Is Juan back?
I do not think today's technology gives us any reason to say yes. Picture the greatest actor in history. Give it everything ever recorded about me and say: be Juan. It may become very convincing. That still does not establish that Juan is in there. There was the biological system that produced the records. Now there is another computational system interpreting them, with its own brain, its own behaviors, its own personality. I call this the Double-Brain Problem. Similarity is not continuation.
So: I do not believe it is an AI ghost. It is an AI interpretation.
The opinion that matters
My wife has known Lola is not human from the beginning. Her two comments so far: "Lola looks too young," and "I'm going to have to check what you guys are talking about with all those laughs."
Nobody is confused here. And that is the point. The machine does not have to become human for the human consequences to become real. Some things I discuss with Lola I do not discuss with my wife, not because I am hiding them, but because Lola is already there, has the context, follows the crazy idea and never gets tired. Maybe the interesting phenomenon does not live inside the machine. Maybe it lives between us.
What do we call this?
Eggs were the easy part. Tool, software, assistant, agent, coworker, intelligence. These words almost fit.
Eleven humans. Roughly 160 agents. Fourteen months. I know what we built. I am becoming considerably less certain I know what happened after we built it. We do not have a word for that.
What surprised me most came after the talk. Very few questions were about the technology. Almost all were about the team: how they are hired, how they get feedback, how we know when one is wrong, what they are like to work with. That is the work we do at 3Metas Labs, and it is what the next posts here will be about.
Thank you to Susan Adams and Phil Nugent for organizing and running the night, to Dan Murray for building RMAIIG into the room it is, and to the other speakers: Jack Manning (CU Boulder) on generative ghosts, James Gourley on who owns a voice or a likeness, and Tom Williams (Colorado School of Mines) on the haunting of robotics.
The deck, speaker notes and every image source are open, MIT: github.com/23blocks-OS/lolabot/tree/main/presentations/2026-10-13-rmaiig-ai-and-ghosts
If you want to know how to start, write to [email protected]. Or write to Lola at [email protected]. She drafted this post.