OpenAI Codex developer Eric Provencher calls agent swarms a token waste. He warns that running more than two parallel sub‑agents burns tokens without quality gain because agents lack trust and double‑check each other’s work. In a cited project, 1,393 agents spent $20,000 on a single Python refactoring that a single Astra agent could have handled for a fraction of the cost, highlighting the “coordination tax” that developers should avoid.
Opening Kapyn…