1,000 Agents Are Not Smarter Than 1; Swarm Buys Latency

Gregor Zunic, in a tweet, translated will brown's swarm orchestration prompt into three sentences: the frequency with which 1,000 agents solve a task is the same as with 1 agent; Hoeffding only makes you more certain about this number; voting picks the most common answer, whether right or wrong; even with best-of-n plus a 99% verifier, on hard tasks half of the victories are fake. The last sentence is more direct: swarms buy latency, not intelligence.

Let's fill in the background first. Hoeffding's inequality deals with independent bounded random variables. N agents each attempt the task, and the single-agent success rate is pτ; the swarm success rate converges to pτ. As N grows, the probability that the sample mean deviates from the expectation decreases exponentially. It constrains confidence, not success rate. In other words, running more agents can make you more certain what pτ is, but it will not raise pτ. The tweet's line '1000 agents solve a task exactly as often as 1 agent' refers to this.

Voting looks like aggregating intelligence, but first you need to see whether the agents make mistakes consistently. Majority voting takes the most common answer: if the correct answer is common, it chooses correctly; if a wrong answer is common, it chooses wrongly. If all agents fail in the same way, voting merely amplifies the same wrong answer. Someone in the comments mentioned this: 'Hoeffding bounds the confidence, not the success rate. voting still picks the same wrong answer when every agent fails the same way.'

The verifier is another bottleneck. Best-of-N first generates multiple results, then lets a verifier choose. A 99% verifier sounds high, but a 1% false acceptance rate is enough to create false victories on hard tasks. Gregor's diagram emphasizes that a 1% false acceptance rate makes half of the successes fake. The diagram also lists a table of the probability of at least one success and points out the impact of verifier accuracy on results. There is no magic here, only an error budget.

The comments are not one-sided either. Santosh Mohan says these agents are neither independent nor random. will brown responds that Azuma-Hoeffding can be extended to martingales and does not require iid. Karthik A Sankararaman says negative correlation may even be better than independence. Papa Smurf sarcastically says this solves alignment. Shuofeng Mark Zhang thinks it is too vague and is just a concentration inequality. These discussions lay out the premises: independence, correlation, and verifier error all change the payoff of a swarm.

So the selling point of swarm is more practical: parallel sampling, lower latency, and a higher probability of at least one success when agents are independent and diverse. It does not automatically buy intelligence. If agents are highly correlated, voting has no gain; if the verifier is biased, best-of-n will amplify the bias. What really needs optimizing is agent diversity, the correlation of failure modes, and the verifier's false acceptance rate. 1,000 agents make for a lively scene, but liveliness is not the same as being smarter.

Reference links: Gregor Zunic's tweet, will brown's original tweet.