There's a quiet fork in the road for AI agents, and almost everyone is taking the wrong branch without noticing.
The open-source world is busy building role agents. "Content Creator." "Sales Rep." "Product Manager." Drop in a system prompt, wire up some tools, and you've got an agent that behaves like a generic version of a job title. It's a sensible idea and it's wildly popular.
I think it's the wrong unit. The future isn't the position. It's the person.
The tell is in the price
Look at what the market is actually doing. The Agency — one of the most popular agent repositories going — has around fifteen thousand GitHub stars and ships sixty-one role agents, free, under an MIT licence.
Now sit with that last number. Sixty-one specialised agents, free. Why free? Because a role is a position, and a position is generic by definition — there's nothing proprietary about "act as a content strategist and give me best practice." Anyone can write it. Everyone can copy it. The price of a thing that anyone can replicate trends to zero, and these already have.
So the fifteen thousand stars aren't proof that role agents are valuable. They're proof of demand — people clearly want specialised agents at scale. They're settling for generic ones because the licensed version doesn't exist yet. That gap between what people are reaching for and what they're stuck with is the whole opportunity.
A role is a position. Anyone can copy it. A person is not. That difference is the entire product.
Generic role versus specific person
Make it concrete. Ask a generic "marketing agent" how to fix your content strategy and you'll get the bland average of everything ever written on the subject — post consistently, know your audience, repurpose across channels. All true. All useless, because it's what everyone already half-knows.
Now ask an agent built from one specific person's actual judgment. You don't get the average. You get positions — the sharp, sometimes contrarian calls that person is actually known for.
| Generic role agent | Person-based agent | |
|---|---|---|
| On SEO | "optimise for search intent" | "SEO is dead for most small businesses — go where attention is" |
| On posting | "be consistent" | "post 7–10 TikToks a day testing hooks" |
| On asking for the sale | "include a clear CTA" | "give value three times before you ask once" |
| What you get | best practice | one person's hard-won judgment |
Those aren't best practices. They're a specific operator's bets, made from twenty years of being right and wrong in public. You might disagree with them — that's the point. A position you can disagree with is worth something. An average you can't even argue with is worth nothing.
What makes a person-based agent actually different
It's not just a spicier prompt. Encoding a person rather than a position takes four things a role agent structurally cannot have.
Core beliefs. Every real expert has a few stable axioms that anchor everything else they say. "Attention is the most valuable asset in business" sits underneath every tactical call the person makes. A person-based agent holds those beliefs as the spine of every answer, so the advice hangs together instead of contradicting itself topic to topic. A role agent has no spine — just whatever the prompt happened to say.
Evolution. People change their minds, and the change is the interesting part. "I used to tell everyone to pick one platform and go deep. Now I say be everywhere. Here's what changed." A role agent has no past to have evolved from. A person-based one tracks the shift and narrates it — turning a contradiction into a lesson.
Boundaries. A good expert tells you when you've wandered outside their lane: "I haven't really worked on supply chain, but my principle about killing friction would point you here." That honesty is the expertise. A generic role agent will confidently answer anything, because it has no sense of where one person's knowledge actually stops.
Unified voice. Ask a person-based agent about hiring, then about pricing, then about content, and it sounds like the same human throughout — same instincts, same turns of phrase, same worldview. A role agent is only ever as coherent as its last instruction.
Why this is the defensible one
Here's the part that matters if you're thinking about moats rather than demos.
That's why "person over position" isn't a slogan I'm fond of — it's a description of what's actually defensible. Role agents commoditise, because positions are generic and the open-source community will always ship a free one. Person-based agents don't, because a specific human's judgment, reconstructed faithfully and kept current, can't be cloned from the outside. You either have the licence and the participation, or you have a guess.
The whole industry is optimising the part that's already free. The valuable question was never "how do I build a content strategist?" It's "whose content strategy judgment do I actually want?" The first has sixty-one free answers. The second has exactly one per person — and that's the market worth building.
Cheers, Adam