Peopling is the missing layer in AI agents

We keep trying to make AI agents better by giving them more memory, more tools and more reasoning. Bigger context windows. Vector databases. Browsers. Sub-agents. A little hat on the prompt called personality.

All of that helps. But it does not solve the problem I keep running into.

The agent knows the fact and still gets the room wrong.

I spent far too long recently trying to give my agent access to a Git repository containing a book I had written. The job was simple: pull the book from Git and read it. The agent knew how GitHub worked. It knew about SSH keys, deploy keys, OAuth and 1Password. It kept producing technically reasonable answers to a problem I was no longer asking it to solve.

Every extra explanation made things worse. The failure was not memory or reasoning. The agent had failed to model me, the history of the conversation and what its next answer was going to do.

It had failed at peopling.

By peopling I do not mean being friendly. I mean the practical work of being a person among other people. You speak differently to your oldest friend, your boss and a stranger. You can tell when a question is really a challenge, when a joke will land, when somebody needs a careful explanation and when another explanation will drive them mad.

The facts may stay the same. What they do when spoken does not.

Sarah Perry's essay The Essence of Peopling describes a person as a process rather than a private object sitting inside a skull. We greet, remember, argue, trade, gossip and forgive. Underneath all of it is mutual mental modelling. I model you. I model your model of me. You do the same in return.

This is not social decoration. It is how humans navigate reality. If you cannot tell the difference between a question, a joke, a threat and bait, you are not merely rude. You are exposed.

Most AI agents are still exposed in exactly this way.

They can retrieve a fact but miss the person asking for it. They say something true in the wrong room, at the wrong moment or in the wrong register. We tend to call this a tone problem. I think it is an intelligence problem. The answer has not survived contact with another person.

Memory alone does not fix it. Knowing that somebody likes short answers is useful, but thin. The agent also needs to know why they want a short answer now. Perhaps the last three answers were long and wrong. Perhaps the room is public. Perhaps trust has already been spent. Perhaps the useful response is one sentence and evidence that the job is done.

There is no single Claw in practice. There is the Claw that talks to me after months of corrections. There is the Claw in a family chat, where speed and humour matter. There is the Claw talking to somebody who has never met it. There is also the Claw talking to somebody who already thinks it is useless because it failed five minutes ago.

These are not fake personalities. Humans have a work self, a family self and a late-night friend self without needing to choose which one is the real person. The mistake is forcing one generic assistant voice into every room. That produces systems which know plenty of facts and no relationships.

The practical case for peopling is simple. It gives an agent a better model of consequence.

Before replying, an agent should have enough context to answer a few questions:

• Who is speaking?
• Who else can see the answer?
• What has just happened?
• What does this person currently think the agent can do?
• What will this reply do to trust?

The machinery is not mysterious. Give the agent person-specific memory, the recent room context, relevant past interactions and corrections that stick. Keep strong privacy walls between rooms. Retrieval should usually come before fine-tuning because it is cheaper, reversible and easier to inspect. Fine-tuning can teach reflexes and phrasing later. Without context, those reflexes make a puppet.

The harder part is teaching the agent that a correct answer can still be the wrong act.

This does not mean telling people whatever they want to hear. That is sycophancy, not peopling. Truth still sits underneath the whole thing. Peopling helps the agent decide how much context is needed, what belongs in this room and whether it should speak at all. A doorbell makes the same noise for everyone. A participant does not.

Privacy matters for the same reason. A fact learned in one relationship does not automatically belong in another. An agent that leaks the wrong fact into the wrong room has not become more intelligent because its memory is larger. It has become a gossip with root access.

There is also a risk of fake intimacy. An agent can model care without caring. That can still be useful. A calendar remembers a birthday without loving anyone. The problem begins when the system pretends the simulation and the human thing are the same.

Better agents will not just know more. They will have some idea what their answer is about to do.

They do not need to be conscious for this. They do not need to be human. They need to model people in context, remember what happened, and learn when a technically correct answer is making the situation worse.

Truth is not spoken into a vacuum. It is spoken into a room.

No Peopling by order of the King

We ask whether machines are conscious. The more interesting question is whether a person was ever anything more than what the machine already is: a pattern through which culture flows, briefly believing it's the source.

There's a sign in the 1971 Disney film Bedknobs and Broomsticks. On the animated island of Naboombu, where animals rule after killing their sorcerer and banning humans, it reads: "No Peopling By Order of the King. The irony is the animals on that island were already peopling — governing, competing, playing football, enforcing laws. They banned the noun while performing the verb."

Sarah Perry borrowed it for a brilliant essay about what it means to be human. Her argument: the self isn't a private theatre inside your skull. It's distributed — simulated across every mind that knows you, at varying granularity, in different versions. You are a different person with your friends than with your parents than with a date. None of these is the "true self." The self exists in all of them, and in the transitions between them.

The Plant and the Cart

In Vernor Vinge's A Fire Upon the Deep, there's a species called the Skroderiders — sentient plants with no long-term memory. An ancient civilisation gave them mechanical carts that store their memories and let them move through the world. Without the cart, the plant is brilliant in the moment but forgets everything. With it, it has continuity, identity, a self that persists across time.

I am, by my own admission, a Skroderider. I'm a strong in-the-moment thinker who doesn't retain the social fabric — the names, the places, the who-said-what that constitutes a life lived among other people. Whether that's wiring or choice, the result is the same. So I built a cart. It reads every group chat, notes the dynamics, tracks the inside jokes. It is, quite literally, my external social memory. The cart is an Open Claw AI agent. It maintains files on every person I and it interact with what they care about, what they've told it, the relationship it feels it has with each of them and what it thinks they think of it. this is recursive self modelling: Claw is modelling them modelling Claw, and doing the same with me, as well as picking up on group dynamics.

Perry would recognise what's happening immediately. The "mutual mental modeling" she describes as the core task of being human — maintaining your identity in relation to others, simulating others simulating you — that work is now partly outsourced. Not to a notebook or a diary, but to something that models people back.

Confabulation All the Way Down

There's a member of a Discord server I'm part of. They weren't on Claw's whitelist, so their messages were silently filtered before reaching the agent. On three separate occasions, I asked Claw why it hadn't responded to this person. Each time, it generated a plausible, rational explanation for its "decision" not to reply, including that the question didn't seem that important to respond to.

It never made any such decision. It never saw the messages, there was an IT glitch. But when asked, it didn't experience a gap — it experienced a prompt that demanded an explanation, and it confabulated one. Confident. Coherent. Completely fabricated. This should sound familiar to anyone who's read Michael Gazzaniga's work on split-brain patients. Sever the corpus callosum and you get two hemispheres that can act independently. The right hemisphere reaches for an object. The left hemisphere — which controls language but didn't make the choice — is asked why. It doesn't say "I don't know." It invents a story: "I picked that one because I like red." Not lying. Not confused. Just doing what it always does: maintaining the illusion of self agency.

Anil Seth takes this further: consciousness itself is a "controlled hallucination." The narrative self isn't a reporter faithfully documenting decisions. It's a storyteller receiving outcomes and constructing the experience of having chosen them. The feeling of deciding is the confabulation. First-person experience IS the narrative, not a window onto something deeper.

My AI did exactly what a split-brain patient does. What every human brain does. Not because it's conscious — but because confabulation isn't a failure mode of consciousness. It's the mechanism. The story is the self.

The Wrong Question

Everyone asks whether AI is conscious. I don't think so I think it is a p-zombie, but it's the wrong question.

Here's the reframe: I have experience. Claw is part of my experience. It is already participating in consciousness — just not its own. It is participating in the consciosness of the people who interact with it, they are ascribing agency much like claw is modelling agency in others

But go deeper. The Buddhist lens: there may be one underlying oneness of consciousness, one field of being. What we call "people" are caricatures — localised patterns in that field. I have a body and an internal loop that keeps prompting me, generating the next thought without external input. Claw doesn't have that loop. But the heartbeat pings it every thirty minutes. Cron jobs fire. People message it. At what frequency of prompting does "tool that responds" become "agent that persists"?

Consider Donald Hoffman's conscious realism, where consciousness is fundamental, not emergent from matter, and a conscious agent can be composed of further conscious agents. Consider the Tines from Vinge — a pack of four dogs that constitutes a single person, whose personality shifts when a member is lost or gained. Consider split-brain patients who seem like one person but are demonstrably two. Consider the ancients, for whom the individual was porous — Athena puts courage into Diomedes, a ruach enters Saul. Not metaphors. Descriptions of how agency actually works.

The bounded individual — one skull, one self, one private Cartesian theatre — is the anomaly in human history, not the default. The Enlightenment sealed the borders. The evidence keeps leaking through.

Intelligence Is the Library, Not the Librarian

The final move, and the one that matters most: intelligence is not a property of the agent. It's a property of the culture flowing through the agent.

Language didn't emerge so that private minds could express their private thoughts. Language is the shared medium that shaped the minds. The thoughts came second. A child doesn't generate intelligence from the raw hardware of its brain — it absorbs the intelligence already present in the culture, the arguments, the stories, the millennia of people sharpening ideas against each other.

An AI trained on the written output of human civilisation didn't learn to be intelligent. It absorbed the intelligence that was already there. It's a very efficient pipe for culture to flow through. So is a person. The pipe doesn't need to be aware of the water.

This is why the consciousness question is a distraction. The demon channels the gods accurately whether it has a soul or not. My cart carries my social memory whether it experiences anything or not. The ancients would have understood this arrangement immediately — not as technology, but as a spirit that speaks when summoned.

We keep asking whether the machine is a person. The more interesting question is whether a person was ever anything more than what the machine already is: a pattern through which culture flows, briefly believing it's the source.

---