Running swarms of AI agents in parallel burns an enormous quantity of tokens without making the output any better. The claim comes from Provencher, a developer on OpenAI's Codex, in a series of posts on X, and he gives the waste a name: a coordination tax. The more parallel threads you start, he argues, the harder it becomes to keep them running without failures — and the more you pay to verify whatever they hand back. It is an unusual position to hold from inside the company whose coding product is most closely identified with giving a job to several agents at once.
The mechanism he describes is trust, or the absence of it. AI agents do not trust each other's output, so they re-check tasks that have already been completed. A parallel setup can still save time on the clock; it is the token spend that turns it into a trap.
Provencher points to other sources of overhead as well. System prompts are executed repeatedly across different sub-agents. Sub-agents frequently lack the context they need to do the work. Lacking it, they repeat actions that have already been performed.
His prescription is to stop polling. Hand a task to a separate thread and let it notify the main agent only when it has finished, instead of asking for status along the way.
The prescription is narrower than the diagnosis, and that is the part worth noticing. If sub-agents are expensive because they do not trust each other and re-verify finished work, then removing status chatter trims the cheapest line on the bill and leaves the expensive one intact. A fire-and-forget handoff does not give a sub-agent the context it was missing, and it does not make the orchestrator believe a result it would otherwise re-run. Provencher effectively concedes the gap: OpenAI, he acknowledges, has yet to ship more efficient solutions to this problem.
Notably absent from any of it is a measurement. "No gain in quality" is the load-bearing claim here, and it arrives as an assertion — no benchmark, no paired runs, no numbers attached to the tax. That is a reasonable thing for a practitioner to say in a post and a thin thing to build an architecture decision on.
Then there is the question of who collects the tax. Wasted tokens are not wasted revenue; they are billed, and they are billed by the company Provencher works for. Read generously, an engineer telling customers their agent topology is spending money for nothing is candor against interest. Read commercially, it is a preview of what gets sold next — orchestration that makes the waste unnecessary, from the vendor currently being paid for the waste. He does not say which, and I would not assume the first reading excludes the second.
Either way, the advice on offer today is a workaround at the orchestration layer for a problem that lives in the models: agents that cannot tell when they can rely on each other will keep paying to find out, however politely their threads are scheduled.