OpenAI Codex developer flags multi-agent swarms as economically inefficient

Provencher doesn't just claim swarms are inefficient; he quantifies a specific failure mode (redundant verification across agents) and names it. The missing context: whether this $20k case was a poorly orchestrated implementation or a structural flaw in the swarm approach itself, and whether his alternative (focused smaller models) is actually being adopted or remains theoretical.

This story is largely disconnected from recent coverage in the space. We have no prior Modelwire reporting on multi-agent architecture trends, cost analyses of orchestration layers, or competing design philosophies in agentic systems. Provencher's critique belongs to the emerging conversation about whether the industry is overbuilding complexity into AI systems, but we lack the archive depth to trace how this argument has evolved or whether other researchers have made similar points.

If OpenAI's next agent release (Astra or successor) ships with fewer than N parallel sub-agents or explicitly documents token efficiency gains versus prior swarm designs, that signals Provencher's critique influenced internal product decisions. If the company publishes a technical post defending multi-agent coordination within 6 months, that suggests the critique landed hard enough to warrant a response.