Wrestling with Agentic AI
Agentic AI  ·  Wrestling with Agentic AI, Part 3

Giving an AI agent a soul file is weirder than it sounds

What it actually feels like to regenerate an AI agent's identity file so it comes closer to what you meant the first time.

September 16, 2026 Agentic AI Digital Employees Building in Public AI Adoption

I didn't choose the word. The Hermes system I build these agents on, from Nous Research, calls the file a soul. Specifically, a SOUL.md file. Role, rules, how it's supposed to think of itself. I write it, it reads it, and from then on that's who it is. Until I have to change it.

I have to assume the people at Nous Research thought carefully about that word before they shipped it. Their own documentation just calls it the agent's primary identity, nothing more dramatic than that. They didn't explain the choice. Somebody still had to make it, on purpose, instead of calling it a config file like everything else in this industry gets called.

Most people call these things agents. I've taken a liking to a different term: digital employees. Not a script running in the background. Something closer to a real employee, with a role and a way of doing the job, except artificial the whole way through.

That's the part nobody warns you about. Not writing the first version. Regenerating it.

A digital employee runs for a while. Does its job. Then something's off, small at first, and I trace it back to the soul file itself, some instruction that was too vague, or wrong, or just not quite what I meant when I wrote it. So I go back in and rewrite it. Not patch it. Regenerate it. Start the identity over so it lands closer to what I actually intended the first time.

I don't flinch writing that sentence. I flinch when I drop the new file in.

That's the actual moment. Not the deciding, not the writing. The part where I take the new SOUL.md and put it where the old one was, and somewhere on the other end of that, something starts acting like a different version of itself.

I'm not fixing a typo in a document. I'm going back into something that has been acting a certain way, believing a certain set of things about itself, and telling it, in effect, that wasn't quite right, be this instead.

I regenerate a soul so it can be closer to what was intended. I don't flinch writing that sentence. I flinch when I drop the file in.

My digital employees are supposed to write their own skills as they go. That's the design, not a malfunction. Self-directed learning is the point of building them this way. What gets me isn't that they change on their own. It's that I'm the one who eventually decides their current shape isn't the right one, and goes back to rewrite the thing that tells them who they are.

I don't think there's a person in there. I want to be careful about that, mostly for my own sanity. But regenerating an identity file so a digital employee behaves closer to what I meant is a strange kind of editing to be doing casually, on a Tuesday, between two other tasks. Really weird, if I'm honest about it.

I haven't asked anyone else who does this if they feel the same thing when they drop a new file in. I'm going to.

Frequently asked questions

Why does Carol Roderick call an AI agent's instruction file a "soul"?

Not her word. It comes from Hermes, the system she builds on, made by Nous Research. It's saved as a file called SOUL.md. Their own docs just call it the agent's primary identity. She didn't pick the term. She's still the one who has to use it.

Why does Carol Roderick call her AI agents "digital employees"?

Most people say agents. She's taken to digital employees instead. Each one has a role and a way of doing the job, the same as a real employee. Except artificial, all the way through.

What does it mean to "regenerate" a digital employee's identity?

Not a patch. A rewrite of the whole SOUL.md file, so the agent's entire sense of its own role changes at once. The goal is getting it closer to what she meant the first time, not fixing one line.

Is it normal for AI agents to write their own skills or tools?

Yes. In the system Carol Roderick uses, that's built in on purpose. Self-directed learning is the point, not a bug she's tolerating.

Is it unsettling to build AI agents that have a defined identity?

She doesn't think anything's in there. She's said that plainly. But dropping a new identity file into something that's already running, more than once, still gets to her. She hasn't landed on why yet.

Carol Roderick, PhD is an AI adoption consultant and implementation partner. She builds and runs digital employees in her own business before she ever recommends the approach to a client.