Centaur Collaboration
January 3, 2026
The term “centaur” comes from chess — teams of humans working with AI that outperformed both pure human players and pure AI systems.
Ethan Mollick’s Centaurs and Cyborgs on the Jagged Frontier explores this dynamic in depth — the distinction between centaur (strategic delegation) and cyborg (fluid integration) styles shaped how I think about these workflows.
This isn’t just a chess phenomenon. It’s a design pattern for knowledge work.
The Insight
Neither pure automation nor pure manual work is optimal. The sweet spot is collaboration — but only when the collaboration is designed well.
Bad human-AI collaboration:
- Human does the thinking, AI does the typing (glorified autocomplete)
- AI does the thinking, human just approves (rubber stamping)
Good human-AI collaboration:
- Each contributes distinct capabilities
- The human maintains agency and understanding
- The AI handles scale, recall, and consistency
- The result is better than either could produce alone
Patterns I’ve Found Useful
1. The Explore-Decide-Execute Loop
AI is excellent at exploring possibilities and generating options. Humans are excellent at making judgment calls and decisions. Design workflows that play to both strengths.
2. Externalized Thinking
Use AI as a thinking partner, not just an executor. Explain your reasoning, ask for pushback, use the conversation to develop ideas.
3. Progressive Disclosure
Start with high-level direction, let AI propose approaches, then dive deeper into the parts that matter. Don’t try to specify everything upfront.
Example: A Simple Centaur Workflow
Here’s a pattern I use frequently:
// Human provides intent, AI expands
const workflow = {
human: "I need a function to parse dates",
ai: generateOptions(3), // AI proposes approaches
human: selectAndRefine(), // Human chooses, AI implements
};
The key is maintaining the loop — neither side operates in isolation.
More patterns emerging. This is a living document.
See also: Cognitive Engineering for the broader framing.