An OpenAI Codex developer says two parallel agents is the limit, after a $19,300 run
Eric Provencher, who works on developer experience for OpenAI's Codex, says that running more than two sub-agents at a time is "almost certainly burning tokens for 0 quality gain". He posted that on September 15, eleven days after Nous Research published the bill for a run that dispatched 1,393 of them.
Agent tooling keeps spreading: Unity now ships official Codex and Claude Code plugins for its engine. Image: Unity
Both sides put numbers on the table, which is rarer in this argument than opinions are.
The coordination tax
Provencher's explanation is short: "Agents don't trust each other enough to avoid double-checking everyone's homework." Pressed on it the same evening, he called the overhead a coordination tax and said that in most cases it is "really hard to keep so many parallel lanes running, without spinning out and burning tokens on excessive verification".
He does not say parallelism never works. He says the lanes have to be split cleanly, and that most people do not split them cleanly.
His own workaround, posted on September 12, is to have one agent plan the work, hand it to a second thread and then end its turn, letting the second thread wake the first when it finishes rather than polling it. He describes it as something he does on his personal plan. This is one named engineer's view, not OpenAI policy.
What 1,393 sub-agents actually bought
Nous Research ran the opposite experiment on its own codebase and published the figures. In a write-up by Teknium, the team says a cleanup started on September 2 ran for about nineteen active hours and "dispatched 1,393 subagents, reaching 218 running at once", with the pull request merged on September 4.
The coordination ran in one Python process on an i7 desktop with 64GB of RAM, while Claude Fable 5.1 handled inference remotely.
The results are specific. One file, gateway/run.py, went from 34,847 lines to 5,512. Non-test Python across the repository fell 34.4%, from 1,063,826 lines to 698,363. Files over 5,000 lines dropped from 37 to six, and functions over 300 lines went from 192 to two.
Nous puts the model cost at about $19,300 for the main run, or roughly $25,000 including follow-up sessions, against its own estimate of $150,000 to $1.8m for a small team doing the work by hand over two months to two years. The post is subtitled "How to get $1.8M of value from $19K of tokens".
Provencher replied to the post directly
On September 16 he answered Teknium in public: "You refactored a large python file and spent $20k in tokens on it. You could have given this to a single astra agent and it would have done it at a fraction of the cost. What were all those sub agents doing?"
Nobody has run that counterfactual. There is no published single-agent attempt at the same repository, no independent measurement of quality on either side, and one codebase with one model is not a benchmark.
What both accounts do settle is the shape of the bill. Nous reports its own figures as a success and still records six files over 5,000 lines left standing, longer import times for some entry points, and coupling the refactor did not resolve.
The practical reading for anyone paying by the token is narrower than either post. Parallelism has a price that scales with how badly the work is divided, and the only people publishing that price so far are the ones who spent it. Tooling is racing ahead of the evidence, with Unity now shipping official Codex and Claude Code plugins for its engine.
