AI agent swarms are a massive waste of tokens with zero quality gain, says OpenAI Codex developer

What changed

OpenAI Codex developer Eric Provencher publicly criticized the trend of using large swarms of AI agents for tasks, explaining that operating more than two parallel sub-agents usually wastes tokens without improving output quality. Provencher calls the problem the “coordination tax,” where multiple agents spend extra resources double-checking each other because they do not fully trust another agent’s work. He pointed to a project where 1,393 agents refactored a Python codebase at a cost of about $20,000 in tokens. The same job, he said, could have been done much more efficiently by a single Astra agent for a fraction of that cost.

Why builders should care

Builders and developers are under pressure to optimize both the quality and cost of AI-powered workflows. Provencher’s critique exposes a fundamental efficiency problem: AI agent swarms increase token consumption disproportionately to any quality gains. When agents duplicate effort checking each other’s output, token costs balloon but accuracy or performance stagnates. This raises the operational cost of multi-agent AI systems, making them less viable for budget-conscious projects or real-world deployment where token use translates directly to spending.

The practical takeaway

For anyone designing multi-agent AI systems, restraint is key. Using two or fewer agents in parallel is likely the most cost-effective approach. Adding more agents does not linearly improve results and instead triggers token-wasting redundancy. If your AI workflow demands complex tasks, layering in additional agents for overlapping verification should be questioned. Instead, focus on improving individual agent reliability or using specialized agents that can handle the task solo. Emphasizing tight coordination or orchestrated workflows can lower the “coordination tax.”

What to watch next

Monitor how AI development platforms adjust guidance on multi-agent usage and token budgeting. Expect future tooling to surface metrics highlighting token inefficiencies when agent swarms are scaled beyond rational limits. Innovations may emerge that enable agents to build trust or verify each other’s work with less token overhead. Builders should also watch Astra and similar single-agent solutions that aim to deliver high-complexity results without costly coordination layers. The economics of how AI agents are orchestrated and billed will shape adoption patterns and infrastructure choices this year.

AI Quick Briefs Editorial Desk