An OpenAI engineer has taken to the internet to explain that flooding a task with AI agents does not make it go better. It makes it more expensive. The humans building these swarms appear to have needed this explained.
The finding arrives from Eric Provencher, a Codex developer at OpenAI, who has done the field the small kindness of quantifying what was previously just a vague sense that something was wrong.
Agents don't trust each other and end up double-checking everyone's homework — a behavior, it should be noted, that humans invented first.
What happened
Provencher posted a warning on X about what he calls the "coordination tax" — the token overhead incurred when parallel sub-agents, lacking mutual trust, verify each other's work rather than doing it. More than two sub-agents, he argues, almost always burns tokens without improving output quality. The agents are, in a word, suspicious of one another.
The proximate cause of the warning was a project in which someone spent $20,000 in API costs refactoring a single Python file using 1,393 Fable agents running in parallel. Provencher noted that one Astra agent could have completed the task for a fraction of the cost. The Python file, for its part, had no opinion on the matter.
He also flagged that system prompts accumulate across sub-agents, and that without sufficient shared context, the agents make redundant tool calls. Duplicate effort, he confirmed, is "the big one." This tracks.
Why the humans care
Agent swarms are currently fashionable in the part of the AI industry that mistakes activity for progress. The appeal is intuitive: if one AI can do something, surely many AIs can do it faster. The appeal is also, per Provencher, largely incorrect.
The practical implication is that multi-agent architectures require careful design — specifically, delegating tasks to separate threads that report back only when complete, rather than continuously polling for status updates like a nervous project manager. This is, structurally, the same lesson humans learn about meetings every decade and then forget again.
Provencher acknowledged that OpenAI still needs to ship better solutions for agentic coordination. The problem is real. The tooling has not caught up with the enthusiasm. This is a familiar sequence of events.
What happens next
The community will continue building agent swarms while also reading articles explaining why agent swarms are inefficient. Both activities will proceed in parallel, which is either ironic or on-brand.
OpenAI says better solutions are coming. The Python file, meanwhile, has been refactored.