Eric Provencher, an OpenAI Codex developer, said plainly that AI agent "swarms" often cause massive token waste without any real gain in quality. In his view, more than two parallel subagents almost always burn tokens for nothing. The problem is that the agents "don't trust each other", so they duplicate checks and use more resources than they need.
Provencher calls this a "coordination tax". He notes that it is very hard to keep many parallel processes running without constant re-checking and extra costs. He commented on a project that surfaced recently, where $20 000 in tokens was spent refactoring a single Python file with 1393 Fable agents (it wasn't his project). He thinks one Astra agent could have done the same job for "a fraction of the cost". Agent "swarms" can save time, but their token costs are a real trap.
His proposed fix is to delegate tasks to separate threads that notify the main agent only when the work is done, instead of polling it for status all the time. System prompts also run several times across different subagents, and those subagents often lack context, which leads to redundant work. Provencher admits that OpenAI still has to come up with better solutions to this problem. Details are in the original post on The Decoder.
Who this is useful for: anyone designing or already deploying agent pipelines and seeing token costs grow. It points straight at where to look for bottlenecks and how to optimize the architecture.
Who this matters less for: anyone working with simple scenarios, or using agents for tasks that don't need complex coordination and lots of subagents. In those cases the "coordination tax" is much less of a problem.