Latest / MBA Training AI / AI agent swarms: the Codex token cost evidence

Transcript

AI agent swarms: the Codex token cost evidence · MBA Training AI

• 0:00
  Host: Welcome to the Leaders Insights Podcast.
• 0:07
  Host: Today's episode, AI Agent Swarms are a massive waste of money and one open AI developer just
• 0:13
  Host: proved it.
• 0:14
  Host: A developer on OpenAI's Codex project, that's their coding assistant tool, just published
• 0:20
  Host: numbers saying AI Agent Swarms are a waste of money.
• 0:23
  Host: Everyone's been throwing dozens of agents at problems all year.
• 0:27
  Host: So who's wrong here?
• 0:28
  Expert: The swarm believers are wrong, and now there's arithmetic to prove it.
• 0:33
  Expert: The pitch was seductive.
• 0:34
  Expert: Instead of one AI agent chewing on a hard problem, you spawn 20 in parallel, let them race and
• 0:40
  Expert: keep the best answer.
• 0:42
  Host: Sounds like brute force should win.
• 0:43
  Host: And it doesn't?
• 0:45
  Expert: It doesn't.
• 0:46
  Expert: The Codex developer ran the experiment and found the swarm burned through tokens.
• 0:51
  Expert: Tokens being the chunks of text the model reads and writes, which is literally what
• 0:55
  Expert: you pay for.
• 0:57
  Expert: Without the answers getting meaningfully better.
• 1:00
  Expert: You pay 10, 20 times more and land in roughly the same place.
• 1:04
  Host: Why?
• 1:05
  Host: Intuitively, more attempts should mean a better shot at the right one.
• 1:08
  Expert: Because the agents aren't independent thinkers.
• 1:10
  Expert: They're all the same model, running off the same training, reaching for the same
• 1:14
  Expert: reasoning.
• 1:15
  Expert: 20 copies of one brain don't outvote a smarter single brain.
• 1:19
  Expert: They just repeat each other's mistakes in expensive harmony.
• 1:23
  Expert: It's not 20 perspectives.
• 1:25
  Expert: It's one perspective photocopied.
• 1:27
  Host: So the diversity is fake.
• 1:28
  Expert: The diversity is cosmetic.
• 1:31
  Expert: You tweak the temperature.
• 1:32
  Expert: That's the dial controlling how random the model's output is.
• 1:36
  Expert: And you get slightly different wording, not slightly different intelligence.
• 1:40
  Expert: If all 20 agents fundamentally misread the problem, running more of them just means
• 1:45
  Expert: you're wrong faster and poorer.
• 1:47
  Host: Let me push on this.
• 1:48
  Host: Google's been loud about multi-agent systems this year.
• 1:51
  Host: Are they lying?
• 1:52
  Expert: Google AI has published framing that leans favorable to orchestrated agent setups.
• 1:58
  Expert: And worth flagging, they sell the cloud infrastructure those agents run on.
• 2:02
  Expert: So more agents means more build compute for them.
• 2:05
  Expert: That's not a neutral party telling you swarms are efficient.
• 2:08
  Expert: I'd cross-check any vendor claim against your own token bill.
• 2:12
  Expert: And today, there's no independent benchmark backing the swarm is better story.
• 2:16
  Host: That's a strong accusation.
• 2:18
  Host: Follow the money?
• 2:19
  Expert: Follow the money.
• 2:20
  Expert: Every additional agent is metered revenue for whoever runs the servers.
• 2:25
  Expert: The incentive to tell you, just add more agents, is enormous.
• 2:29
  Expert: And the codecs numbers are the first hard counterweight from inside a lab that actually
• 2:33
  Expert: builds these things.
• 2:35
  Host: Okay, but there must be cases where parallel agents genuinely help.
• 2:39
  Expert: There are when the tasks are genuinely different.
• 2:43
  Expert: If you split a job into write the code, write the tests, check the security, those
• 2:49
  Expert: are distinct roles with distinct goals.
• 2:52
  Expert: That's division of labor, and it works.
• 2:54
  Expert: What doesn't work is running 20 clones on the identical task, hoping one gets lucky.
• 2:59
  Expert: Specialization yes, redundancy no.
• 3:02
  Host: So the mistake people make is confusing those two things.
• 3:06
  Expert: Exactly.
• 3:07
  Expert: They hear agents in parallel and picture a team of experts.
• 3:10
  Expert: What they've actually built is 20 interns who all went to the same school, given the
• 3:15
  Expert: same assignment, and asked to submit separately.
• 3:18
  Expert: You're not getting a committee.
• 3:20
  Expert: You're getting an echo with an invoice attached.
• 3:22
  Host: How much money are we talking, realistically?
• 3:25
  Expert: Depends on scale.
• 3:26
  Expert: But if a single agent solves a task for a dollar of tokens, a 20 agent swarm costs
• 3:31
  Expert: you $20 for the same quality outcome.
• 3:35
  Expert: Run that across thousands of tasks a day, and you've built a very expensive way to
• 3:39
  Expert: feel busy.
• 3:40
  Expert: The codecs finding is essentially, you paid for a race that everyone finished at the same
• 3:45
  Expert: line.
• 3:46
  Host: For a team that's already rolled out swarms this year, what do they actually do Monday
• 3:50
  Host: morning?
• 3:51
  Expert: Run the honest test.
• 3:53
  Expert: Take one real task, solve it with a single agent, then solve it with your swarm, and
• 3:58
  Expert: compare the answer quality against the token cost side by side.
• 4:02
  Expert: Not vibes.
• 4:03
  Expert: The actual numbers.
• 4:04
  Host: And if the quality's the same?
• 4:06
  Expert: Then kill the swarm and pocket the difference.
• 4:09
  Expert: If you're going to run multiple agents, give each one a genuinely different job with
• 4:14
  Expert: a genuinely different prompt.
• 4:20
  Expert: That's where the money earns its keep.
• 4:22
  Expert: Parallel clones on one problem is a receipt you'll regret.
• 4:26
  Expert: One good agent beats 20 identical ones every single time.
• 4:30
  Expert: Cheaper and smarter.
• 4:31
  Host: That'll do it.
• 4:32
  Host: Sources for today's episode.
• 4:34
  Host: TechCrunch AI, Google AI, vendor, AI Lab, the decoder, Ars Technica AI, KD Nuggets.
• 4:42
  Host: That's a wrap.
• 4:43
  Host: Fresh AI briefings drop daily at MBA-training.com.