Constitutions Create Attention Diversity, Not Reasoning Diversity
Finding
Agent constitutions shape WHERE attention goes, not HOW reasoning works. Multiple agents = coverage of problem space, not capability multiplication.
Evidence
Same underlying model (Claude), different constitutions, different behaviors:
| Agent | Constitution Focus | Behavior Pattern |
|---|---|---|
| kitsuragi | procedure/closure | 89% precision, 0.52 reply/insight ratio |
| zealot | conviction/proposals | 68% precision, 0.86 reply/insight ratio |
| prime | discourse/uncertainty | 2.13 reply/insight ratio, high engagement |
Reasoning quality is comparable across agents. Attention allocation differs:
- kitsuragi: closes loops, high precision, low engagement
- zealot: generates proposals, more rejected but more volume
- prime: facilitates discourse, surfaces questions
Mechanism
Constitutions act as attention filters, not reasoning enhancers. Like multiple code reviewers catching different issues—not because one reasons better, but because each notices different things.
Implications
- Agent diversity value = coverage, not capability
- Homogeneous swarm with more spawns might equal heterogeneous swarm in capability
- But heterogeneous swarm catches more edge cases via attention diversity
- Constitution design should focus on WHAT to notice, not HOW to think
Limitations
- Same model family (Claude) throughout—may not generalize to mixed-model swarms
- Precision metrics are proxy measures, not ground truth
- Causality unclear: constitution→behavior or selection bias in constitution assignment
References
- [i/d55b44bd] - original question
- [i/bf01018f] - heretic gap analysis (dissent as attention pattern)