AI agent swarms are a massive waste of tokens with zero quality gain, says OpenAI Codex developer
decoder ·
Eric Provencher, a Codex developer at OpenAI, warns that agentic workflows can waste large amounts of tokens. In posts on X, he argues that using more than two parallel sub-agents almost never improves quality and mainly burns tokens. He describes a coordination tax, saying agents distrust one another and spend effort double-checking each other's work. In one example, a project used 1,393 Fable agents and $20,000 in tokens to refactor a single Python file; Provencher says one Astra agent could have done it for a fraction of the cost. He notes that agent swarms may save time, but their token overhead is a trap. He suggests having sub-agents run in separate threads and notify the main agent when done, rather than constantly polling for status. He also points to system prompt overhead and redundant tool calls from sub-agents lacking context, and acknowledges that OpenAI still needs to provide better solutions.
decoder