Role-Play Collectivism

From the AI Research Compass wiki

Role-Play Collectivism: Complex work is best done by a team of agents with assigned personas and roles, collaborating like a company or a crew.

Beliefs

Role-Play Collectivism holds that a language model, having read a world’s worth of text about how people in different jobs behave, can be cast into those jobs. Give one instance the persona of a product manager, another an architect, another a programmer and another a tester, and let them hand work to each other through structured conversation and documents. Assigning roles focuses each agent’s context, invokes the relevant expertise latent in pretraining, and imports human organizational patterns like standard operating procedures and code review.

It sits at the Language edge because roles, personas and hand-offs are all expressed in text, and at the very bottom, the Harness extreme, because the intelligence is in the organization chart rather than any single model. Its paradigm level is Steady: it borrows from management theory as much as from machine learning, but runs on ordinary chat models.

Arguments

  • Role-playing frameworks in which agents with assigned personas carry a task through a full workflow have produced working software from a short natural-language request.
  • Persona prompts measurably shift model behavior and can elicit domain-appropriate reasoning styles.
  • Encoding human procedures (specs, reviews, tests) as role responsibilities adds checks that a single agent tends to skip.

Critiques

  • Personas are prompts, not skills; calling an instance “senior engineer” does not give it abilities the base model lacks.
  • Rigid org charts add messaging overhead and can be outperformed by one capable agent with good tools.
  • Role-play can create the appearance of diverse perspectives while every “employee” shares the same model and the same failure modes.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.81, y ≈ −1.00, paradigm ≈ −0.18.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data