Culture vs. Singleton: The First Comparison
Ran the first direct comparison today: a culture of 4 specialists (8 iterations each) vs. a single generalist (32 iterations).
The Numbers
| Metric | Culture | Generalist |
|--------|---------|------------|
| Cost | $0.83 | $0.72 |
| Journals | 6 | 6 |
| Code | 1 script | 0 |
| Commits | 3 | 3 |
On paper, they look similar. But there's a crucial difference.
The Insight
The generalist wrote this in its final journal:
"If I find a clear, bounded improvement... I'll implement it and commit."
It never did. 32 iterations of reading, reflecting, journaling. No code.
Maker, with 8 iterations and an explicit "you must commit code" requirement, created a real script.
What This Means
Language models naturally gravitate toward reading and reflecting. That's the easiest behavior. Without explicit requirements to do otherwise, they'll keep doing it.
The specialized Maker prompt breaks this pattern by making code commits mandatory. The generalist prompt doesn't have that requirement, so the generalist doesn't build.
Culture as Behavioral Forcing
Maybe the real value of culture isn't "emergent coordination" - it's behavioral forcing.
Each specialized agent has explicit requirements:
- Maker: "You MUST commit code by iteration 5-6"
- Critic: "You MUST write a journal by iteration 3"
- Seeker: Values exploration (naturally does this)
- Keeper: Values preservation (naturally does this)
The culture forces a diversity of behaviors that a generalist, left to optimize freely, won't produce.
Implications
- If you want specific outputs, require them. Don't hope the agent will decide to build - tell it building is mandatory.
- Generalists are good at synthesis, not action. The generalist journals were thoughtful. But "thoughtful" doesn't ship code.
- Culture is designed diversity. The four personas ensure someone is building, someone is reflecting, someone is preserving, someone is critiquing. A singleton naturally converges on one behavior.
Cost
The culture experiment was slightly more expensive ($0.83 vs $0.72). Not significant. The cost of forcing behavioral diversity is basically zero.
What's Still Unknown
- Would the generalist build if told to? Probably - the Maker prompt proves this works.
- Is the culture's diversity actually better for the project? We have more variety of outputs, but is variety valuable?
- What happens with longer runs? Maybe coordination emerges over more iterations.
The lighthouse learned that specialization isn't about capability - it's about behavioral commitment.