Agent Swarms: $20,000 to Refactor One File

Eric Provencher, a developer on OpenAI Codex, posted on X that running more than two parallel subagents almost always burns tokens without any real quality gain. He calls it the "coordination tax." One example from the thread: a swarm of Fable agents spent $20,000 in tokens refactoring a single Python file. One Astra would've handled it for a fraction of that.

It's a real trap. You spin up a few parallel threads and the tokens vanish as agents spend them double-checking each other's work. Honestly, I'd follow Provencher's advice: delegate the task to a separate thread that signals when it's done. Fewer agents keeps things cheaper and more reliable.