Agent Swarms: $20,000 for a single file refactor

As OpenAI Codex developer Eric Provencher points out on X, more than 2 parallel subagents tend to burn through tokens with little added value. This is what he calls a “coordination tax”. For example, in this thread, swarms of Fable agents burned through $20,000 of tokens by collectively refactoring a Python file. The same task could have been accomplished far more efficiently and cheaply by an Astra.

This can become a vicious cycle. Start running some parallel threads and before you know it, all your tokens are spent as agents verify each other’s output. Personally, I think Provencher has the right idea here — fork off the task to a separate thread and let it ping you when it’s done. It will be more token efficient and robust.