AI agent swarms: the Codex token cost evidence

September 17, 2026 · 5 min · Free, with full transcript

Does running twenty AI agents in parallel on the same problem produce better answers? A developer on OpenAI's Codex project ran the experiment and found swarms burn 10 to 20 times more tokens while landing in roughly the same place. The argument here is that the agents are not independent thinkers: they share one model, one training set and one reasoning pattern, so temperature changes give different wording rather than different intelligence. You come away with a test you can run Monday morning, a way to read vendor claims from Google AI and other cloud providers against your own token bill, and a clear line between specialization, which works, and redundancy, which does not.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Run one real task through a single agent and through your swarm, then compare answer quality against token cost using actual numbers rather than impressions.
  • If swarm quality matches single agent quality, shut the swarm down and keep the token savings.
  • Give each parallel agent a genuinely different job and prompt, such as write the code, write the tests, check the security.
  • Treat vendor claims about multi-agent orchestration with caution, since cloud providers bill for every extra agent you run.
  • Remember that raising temperature varies wording, not reasoning, so extra clones repeat the same misreading of the problem.
Read the full transcript

Host:Welcome to the Leaders Insights Podcast. Today's episode, AI Agent Swarms are a massive waste of money and one open AI developer just proved it. A developer on OpenAI's Codex project, that's their coding assistant tool, just published numbers saying AI Agent Swarms are a waste of money. Everyone's been throwing dozens of agents at problems all year. So who's wrong here?

Expert:The swarm believers are wrong, and now there's arithmetic to prove it. The pitch was seductive. Instead of one AI agent chewing on a hard problem, you spawn 20 in parallel, let them race and keep the best answer.

Host:Sounds like brute force should win. And it doesn't?

Expert:It doesn't. The Codex developer ran the experiment and found the swarm burned through tokens. Tokens being the chunks of text the model reads and writes, which is literally what you pay for. Without the answers getting meaningfully better. You pay 10, 20 times more and land in roughly the same place.

Host:Why? Intuitively, more attempts should mean a better shot at the right one.

Expert:Because the agents aren't independent thinkers. They're all the same model, running off the same training, reaching for the same reasoning. 20 copies of one brain don't outvote a smarter single brain. They just repeat each other's mistakes in expensive harmony. It's not 20 perspectives. It's one perspective photocopied.

Host:So the diversity is fake.

Expert:The diversity is cosmetic. You tweak the temperature. That's the dial controlling how random the model's output is. And you get slightly different wording, not slightly different intelligence. If all 20 agents fundamentally misread the problem, running more of them just means you're wrong faster and poorer.

Host:Let me push on this. Google's been loud about multi-agent systems this year. Are they lying?

Expert:Google AI has published framing that leans favorable to orchestrated agent setups. And worth flagging, they sell the cloud infrastructure those agents run on. So more agents means more build compute for them. That's not a neutral party telling you swarms are efficient. I'd cross-check any vendor claim against your own token bill. And today, there's no independent benchmark backing the swarm is better story.

Host:That's a strong accusation. Follow the money?

Expert:Follow the money. Every additional agent is metered revenue for whoever runs the servers. The incentive to tell you, just add more agents, is enormous. And the codecs numbers are the first hard counterweight from inside a lab that actually builds these things.

Host:Okay, but there must be cases where parallel agents genuinely help.

Expert:There are when the tasks are genuinely different. If you split a job into write the code, write the tests, check the security, those are distinct roles with distinct goals. That's division of labor, and it works. What doesn't work is running 20 clones on the identical task, hoping one gets lucky. Specialization yes, redundancy no.

Host:So the mistake people make is confusing those two things.

Expert:Exactly. They hear agents in parallel and picture a team of experts. What they've actually built is 20 interns who all went to the same school, given the same assignment, and asked to submit separately. You're not getting a committee. You're getting an echo with an invoice attached.

Host:How much money are we talking, realistically?

Expert:Depends on scale. But if a single agent solves a task for a dollar of tokens, a 20 agent swarm costs you $20 for the same quality outcome. Run that across thousands of tasks a day, and you've built a very expensive way to feel busy. The codecs finding is essentially, you paid for a race that everyone finished at the same line.

Host:For a team that's already rolled out swarms this year, what do they actually do Monday morning?

Expert:Run the honest test. Take one real task, solve it with a single agent, then solve it with your swarm, and compare the answer quality against the token cost side by side. Not vibes. The actual numbers.

Host:And if the quality's the same?

Expert:Then kill the swarm and pocket the difference. If you're going to run multiple agents, give each one a genuinely different job with a genuinely different prompt. That's where the money earns its keep. Parallel clones on one problem is a receipt you'll regret. One good agent beats 20 identical ones every single time. Cheaper and smarter.

Host:That'll do it. Sources for today's episode. TechCrunch AI, Google AI, vendor, AI Lab, the decoder, Ars Technica AI, KD Nuggets. That's a wrap. Fresh AI briefings drop daily at MBA-training.com.

Read and go deeper