OpenAI Codex Developers Critique Token Inefficiency in AI Agent Swarms

Matthias Bastian, a developer at OpenAI Codex, stated that "AI agent swarms are a massive waste of tokens with zero quality gain." In a recent project, 1,393 Fable agents were utilized, leading to a token expenditure of $20,000 for refactoring a single Python file. Eric Provencher, another Codex developer, emphasized that using more than two parallel sub-agents typically results in token waste without enhancing output quality. He introduced the term "coordination tax," explaining that managing multiple agents complicates processes and leads to excessive verification efforts, which further drains tokens. Provencher noted that a single Astra agent could have accomplished the same task at a fraction of the cost, pointing out that the token overhead is a significant trap for developers. He identified duplicate efforts as a major issue, asserting that OpenAI still needs to ship better solutions to address these inefficiencies. This critique follows OpenAI's ongoing exploration of optimizing AI tools, as seen in their recent launches and workshops aimed at improving AI applications.