Centaur Work explains how human‑AI control over direction (WHAT) and execution (HOW) creates centaur, dynamic, and self‑automator modes.
The head steers, the legs run. The only question is how far down your control reaches.
What It Is
Human and AI as one being, with a clear division of labor. The centaur is the image: the human on top (steering), the AI legs below (carrying). The seam between them is the boundary of your control, and it can sit higher or lower.
So working with AI isn't on-or-off, it's a question of how far your control reaches down the body. A research group around Dell'Acqua and Mollick (Harvard, 2025, the same group behind the Jagged Frontier) measured exactly that: 244 BCG consultants, 4,975 interactions, the same task. They found three modes, depending on who decides the WHAT (the direction) and the HOW (the execution).
Three Modes, One Question: how far does your control reach?
HOW decided by human | HOW decided by AI | |
WHAT decided by human | Centaur: you lead WHAT and HOW, AI as a precise tool. 14%, highest quality. | Cyborg: you set the direction, the AI finds the way. 60%, dynamic mode. |
WHAT decided by AI | empty: doesn't occur | Self-Automator: AI decides WHAT and HOW. 27%. |
Three cells are occupied, one stays empty: whoever gives up the WHAT almost always gives up the HOW too. There is no stable middle ground of "the AI decides what, I decide how".
Only 14% were true centaurs, and they produced the most accurate work. Cyborg isn't a failure mode: a fluid back-and-forth with the AI can be exactly right for the task. Self-Automator becomes the problem when you hand over the direction too.
How To Spot It
- Someone gives the AI a clear direction and checks its output deliberately, instead of just adopting it
- On dynamic tasks a fluid back-and-forth emerges without a clean handoff (Cyborg), and the judgment still stays with the human
- The opposite: "just let the AI do it", without checking the direction or the result
What To Do (FL1 - Team Level)
- Set the seam deliberately: decide per task how much of the HOW you hand over. That's a design choice, not a character trait
- Keep the WHAT: holding the direction in your own hands is what decides the learning effect, not the amount of delegation
- Use the AI as a teacher, not as cheap labor: keep the WHAT and you keep learning (Org Topologies calls this Multi-Learning). Hand over the direction and the learning stalls
- Align to the frontier: deep inside the Jagged Frontier delegating can be sensible, near the edge you need the human on top
- Distinguish it from directing: actively directing many agents (Agent Boss) or a founder strategically orchestrating (One-Person Unicorn) keeps the WHAT. That's the opposite of abdication, not the same thing
The Trap
Handing over the WHAT without thinking. Self-automation isn't inherently bad, inside the frontier it can be efficient. The trap is the thoughtless release of the direction: fast, polished, but without depth and without learning. Scaled up across the whole organization, it stays stuck in old patterns instead of getting smarter with AI (First, Org Design).