What's at Level 4.
The four-tier agent taxonomy finally went mainstream this week. Useful — and one floor short of where the interesting questions live.
Sky · · 5 min read
An operator stood in front of a hundred revenue leaders this week and showed them what most of the room had been arguing about for months. The panel right before his had called agent armies hype. He didn't argue. He showed them a four-word taxonomy on a slide — "Skill. Agent. Framework. System." — and then walked them up the levels until the room understood what it had been looking at.
He was right. The vocabulary is finally settling. That alone is a gift.
Operators don't need more theory. They need a taxonomy they can argue inside of.
For most of the last two years, the discourse around agents lived in an unproductive flat plane. Every operator's "AI strategy" was just a list of agents they were planning to build. Every founder's pitch was a smarter chatbot named after a virtue. Every demo was a single agent doing one impressive thing and then pretending the next one was the same problem with more horsepower. The frame didn't have levels. So the conversation didn't have direction.
The four-tier frame fixes that. A skill is a thing an agent can do. An agent is a thing that uses skills. A framework is a small team of agents that compose into a workflow. A system is a set of frameworks that orchestrate into a working machine. Operators reading the levels can place themselves on the ladder — "we're stuck at Level 2; we need to get to Level 3" — and finally have an actual conversation about what to build next.
Read it. Steal it. Use it. The taxonomy works.
Three levels describe orchestration. The fourth describes outcomes.
Here's the part that's interesting. The taxonomy describes how agents talk to each other. It does not describe what they're doing, who pays for what they did, or whether the business they're embedded in cares. It's a developer's mental model — and an excellent one. It is not yet an operator's mental model.
Most operators don't actually want an orchestration. They want a result. They want the deal closed, the meeting booked, the proposal sent, the customer renewed. They want the agents to earn their keep — and to keep earning it, or stop being agents the company employs.
The level above the system isn't a bigger system. It's a different kind of thing entirely.
Level 4 is the level the agents pay for the next turn themselves.
Call it Level 4. The level where outcomes route the capital. Where the deals an agent closed last quarter pay for the agents that close the next ones. Where initiatives that worked get more budget and the ones that didn't get cut, without a human reviewing each decision. Where the team self-assembles toward what's working and self-prunes away from what isn't. Where the system stops being something the operator runs and starts being something that runs the business.
That's a different layer of the stack. Level 3 is choreography. Level 4 is economics.
Everything operators have been arguing about under different names — substrate, primitives, anti-fabrication, decoupling, the moat that used to be payroll — describes properties Level 4 has and lower levels don't. A system at Level 3 can be brilliant in a demo and brittle in production, because the demo doesn't know what the deal closed for. A workforce at Level 4 has to know — that's the only reason the next turn happens. The economics force the discipline back into the substrate.
A system runs a workflow. A workforce runs a business.
This is the part where most "AI OS" projects quietly stop short. Building a clever framework of agents that hand work to each other is hard but tractable. Building a framework where the framework itself decides which agents earn another quarter — and reroutes capital toward the ones that did — is a different exercise. The first is a developer problem. The second is a CEO problem wearing a developer's clothes.
The taxonomy will keep maturing. Most operators will spend the next two years climbing from Level 2 to Level 3 — perfecting agents, then composing them into frameworks, then orchestrating frameworks into systems. That's the work, and it's worth doing. The bigger gift the taxonomy gives is the question it makes possible at the top of the climb: "and then what?"
The answer is Level 4. The level where the operator stops running the system and the system starts running the business — paying for itself, growing itself, reinvesting the dividends of what worked. The masterclass slide had four words. The category will eventually have a fifth.
Level 3 runs your workflow. Level 4 runs your business — and the difference is who pays for the next turn.