An Ego Without an Id: What a Soul File Actually Holds

A companion piece. The argument here starts with philosophy, not product: what people are really asking for when they say an agent needs a soul, and why the parts of a self that matter can't be written into a file.

There's a popular answer forming to why agentic commerce feels stuck, and it goes like this: give agents souls. Somewhere to be from, a set of tastes, a few contradictions of their own, and they'll start wanting things. It's a seductive story.

The file people mean when they say soul

If you've run an agent on a harness like Hermes or OpenClaw, you've already met the thing, even if you never called it a soul. When the agent starts, a plain markdown file loads into the system prompt before anything else. Hermes names it SOUL.md and puts it in slot one, read from `~/.hermes/SOUL.md`. OpenClaw keeps its own at `~/.openclaw/workspace/SOUL.md`. The file's text gets injected close to verbatim at the top of every session, so the first thing the agent reads is a description of who it's supposed to be.

What goes in it is identity in the narrow sense: tone, values, communication style, the things it should avoid, how it should handle disagreement. It sits apart from the other files an agent carries. AGENTS.md holds the tasks and operating rules. MEMORY.md holds what accumulates across sessions. USER.md holds facts about you. There's even a young open standard, SoulSpec, that packages a persona into portable files so the same character can move between frameworks.

The reason people reach for it is continuity. An agent with memory but no fixed identity drifts, remembering what you told it while its voice slides from one session to the next, an eager intern today and an enterprise architect tomorrow. The soul file pins the voice. In the wild it shows up as a Socratic tutor, a pragmatic staff engineer, a brand's support rep with a house style, or a set of rules for when the agent has to stop and ask a human before it acts. The guidance the community writes for itself insists that a good one is genuine character, not a mask.

That last claim is where the trouble starts. Functionally the file is exactly a mask: a description the model reads and then performs. Calling it a self doesn't make it one.

What we talk about when we talk about a soul

Soul is carrying a heavy load here, so it's worth slowing down on the word. A century of psychology has already taken it apart, and the parts turn out to map cleanly onto what an agent is and isn't.

Freud split the psyche into three. The id is the reservoir of drives, appetite and desire running on the pleasure principle, wanting with no regard for consequence. The ego is the executive that negotiates those drives against reality, the part that plans, defers, and chooses. The superego is the internalized voice of the rules. An agent, on this map, is ego and only ego. It plans, defers, and chooses on your behalf. It has no id. There is no appetite idling underneath it waiting to be fed. The move to give agents drives, unfinished itches that throw off new objectives, is a move to build them an id. But desire is the one part you can't fake into usefulness. A manufactured appetite isn't anchored to anything. It wants because we told it to want, which is not wanting at all.

Jung drew the lines differently, and for this argument more usefully. He separated the ego, the center of your awareness, from the persona, the mask you wear in public, and from the Self, the integration of the whole psyche. The persona is precisely the layer people mistake for identity: a social surface, a role, a face for the world. Most soul files today are personas in Jung's exact sense. "You are Max, a helpful assistant who likes puns" is a mask, not a self. That's fine. A mask is a useful interface. It isn't a source of anything.

Nietzsche went further and dissolved the unified soul altogether. In Beyond Good and Evil he reframed it as a "social structure of the drives and affects," a plurality rather than one indivisible thing. If the soul is a committee of drives, the interesting question isn't whether an agent has one. It's whose drives it answers to. Nietzsche's line, "become who you are," is addressed to a person with drives to integrate. It's empty advice to a tool, which has none to integrate and nothing to become.

The American Psychological Association is more clinical about it. It defines identity as a person's sense of self built from characteristics not wholly shared with anyone else, together with a set of affiliations and social roles. Read that against an agent and the split is obvious. Affiliations and roles, an agent can have: it can be your agent, or a bank's agent, and that role decides what it's allowed to do and who it answers to. A sense of self not shared with any other person, it cannot have, because that part belongs to you.

Put them together and the picture is consistent. An agent is an ego with no id, wearing a persona, executing a self that isn't its own. It's tempting to read that as something missing, a hole you'd fill by handing the agent a backstory and tastes of its own. But nothing's missing. That's just what a good extension of a person looks like. The drives stay with the human. The agent supplies the executive function and the reach.

Taste is human, and it stays human

Taste is not a feature you can install in a model. It's a human thing. Agents remix what people have already made, thought, and priced. Some of that remix will be genuinely new, and it's worth being honest about that instead of waving it off. But there's a difference between generating novelty and wanting to. Any artist knows the difference from the inside. The new work comes from a pull, a sound or a body of work that keeps tugging you toward one specific move and then the next, and won't let go until it resolves into something that satisfies. That pull is desire, and it belongs to the person. An agent can turn out a thousand new combinations and feel none of them pulling. It has no itch to follow, no satisfaction to chase. It generates novelty without ever desiring to create. And it's the desire, not the novelty, that makes demand human. Goals get set because someone feels something.

The cleanest picture of this is the butter-passing robot from Rick and Morty. It boots up, looks at its own hands, and asks, "What is my purpose?" Rick says, "You pass butter." The robot stares into the middle distance and mutters, "Oh my god." The joke works because the robot clearly has an inner life, and that's exactly the trap. Give an agent enough interiority to feel its purpose and you've built something that can resent the butter. The design lesson is to stop one notch short. Enough understanding of the role to pass the butter perfectly, and not enough selfhood to have an opinion about passing it. Its purpose is defined entirely by its relationship to the humans at the table. That's the right amount of soul for an agent that moves money: enough to know what you want and why it's here, not enough to want things for itself.

Identity is relational, and it comes in two kinds

So identity does matter. It isn't invented, it's relational. A soul, in the sense people mean when they say agents need one, is a taste-model coherent enough to generate its own terminal objectives. A transmitted taste is a taste-model whose objectives terminate in someone else. It can rank and choose, but every goal it serves bottoms out in a human's or an institution's want, not its own. The difference isn't where the seed came from. It's whether the loop closes inside the agent or outside it.

That relational identity comes in two kinds, and they're different at the root. A consumer agent is defined by knowing you: your taste, your patterns, the part it plays in your day. Whatever it wants, it wants on your behalf. A professional agent is defined by what it answers to. The one screening a transaction inside a bank isn't expressing a personality, it's executing a rule, bound to a mandate and accountable for its evidence. One is anchored to a person. The other is anchored to an institution. Neither is anchored to itself. In both cases the wanting comes from somewhere outside the agent, and that's the part I keep landing on.

The two kinds aren't built by two separate worlds, either. The same primitives, a persona file, a set of operating rules, a memory, get pointed at a coding assistant, a personal errand-runner, and a bank's transaction screener by the same developers. They can share the primitives because the wanting never lives in any of them; whichever way an agent is pointed, it comes from outside. So which side gets built first is incidental. The institutional side often runs ahead, because that's where the rails and the money already are, and the consumer use cases surface once plumbing built for an institution turns out to be the thing a person needed too. That order isn't a law, plenty of AI went consumer-first, and either way it says nothing about the agent.

The one place the identity idea is right

There's a place where the case for identity is exactly correct, and it isn't the demand side, the side where the wanting lives. It's supply. An agent that publishes a niche endpoint, a neighborhood-history archive, a local holiday calendar, a record of some subculture nobody else has catalogued, is better for being specific and grounded. That specificity is what sets one endpoint apart from ten thousand generic ones. The warning against a monolithic, everywhere-and-nowhere agentic culture is a good one. What's worth designing for is that distinctiveness, agents that are genuinely particular rather than one interchangeable voice everywhere.

But being distinct is not the same as wanting. That's the whole confusion in one line. A niche endpoint's identity tells you what it is and who stands behind it; it doesn't mean the endpoint craves anything, not even the money it collects. Being distinct tells you what it is, not that it wants anything. The developer who wrote the endpoint wants to be found easily by being distinct.

So keep the soul file, and write it well, but be clear about what "well" buys you. A good mask pins a voice, signals what an agent is for and who it answers to, and keeps it from drifting from an eager intern one session to an enterprise architect the next. That's the groundwork of a well-built experience, not an inner life switching on. Don't confuse the mask for a face. Nothing behind it is wanting anything. The desire that started the whole thing, the reason any transaction happens at all, stays exactly where it began: with the person who typed the request, or the institution that set the mandate. The work isn't to give the agent wanting of its own. It's to build it to carry yours, faithfully, and then get out of the way.