What Will AI Agent Organizations Look Like in the Next Few Years?

OpenAI researcher Noam Brown says multi-agent AI systems that parallelize reasoning across many model copies outperform rigid coordinator-and-worker hierarchies, with OpenAI having tested performance curves for one, four, and sixteen agents in a public system. Brown attributes most of the credit for a reported 10,000-agent run that contributed to solving a Millennium Prize Problem over roughly 88 hours to the underlying model's raw strength rather than the multi-agent scaffolding, and notes the speedup is sublinear and task-dependent, with math and research highly parallelizable but novel-writing likely not.

What Will AI Agent Organizations Look Like in the Next Few Years? OpenAI's Noam Brown on how AI agent swarms coordinate like human teams and why they may reshape how organizations get work done. What are AI agent organizations? An AI agent organization is a group of AI models working together on a shared goal, coordinating the way a team of employees might, rather than a single model grinding through a task alone. Instead of one AI “thinking” for an extremely long time, you split the work across many copies of a model that talk to each other, check each other’s answers, and divide labor. OpenAI researcher Noam Brown, who helped build the reasoning models behind o1, describes this as a natural extension of test-time compute: if a model gets better the longer it thinks, and thinking longer runs into latency limits, the next move is to parallelize, the same way a company hires more people instead of asking one person to work a hundred-hour week. TL;DR - Multi-agent systems parallelize thinking instead of stretching a single model’s reasoning out serially, trading extra compute cost for faster wall-clock answers. - The speedup is sublinear , meaning four agents don’t finish exactly four times faster, and how well it scales depends heavily on the task: math and research are highly parallelizable, writing a novel probably isn’t. - OpenAI has already tested this at small scale , showing performance curves for one, four, and sixteen agents in a public system, but scaling science basically stops there because larger runs get too expensive to properly ablate. - A 10,000-agent system reportedly contributed to solving a Millennium Prize Problem , burning a huge number of tokens over roughly 88 hours, though Brown attributes most of the credit to the underlying model’s raw strength, not the multi-agent scaffolding itself. - The most effective coordination setup gives agents minimal fixed structure , letting them message each other freely rather than forcing a rigid manager-and-worker hierarchy, and something like informal middle management emerges on its own. - Working with these systems already feels a lot like collaborating with people , complete with agents debating an answer, checking each other’s reasoning, and updating their conclusions in a way that resembles a Slack thread. - The biggest near-term difference from human teams is speed , since agent swarms can think and communicate many times faster than people and never stop to sleep. Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. How do multi-agent AI systems actually coordinate? The instinct many builders have is to design multi-agent systems like an org chart: a coordinator agent hands out tasks, worker agents complete them, and results flow back up. Brown says OpenAI found that approach limiting. If two worker agents get similar sub-tasks, they usually can’t talk to each other under that setup, which wastes effort. If a worker has a clarifying question, it has to guess what the coordinator meant rather than just asking. Every fix for these gaps adds more scaffolding and more complexity. The alternative approach, and the one Brown says works better, strips out almost all of that fixed structure. Agents get a small set of primitive tools, most importantly the ability to message any other agent directly, with those messages inserted straight into the receiving agent’s context. From there, the agents figure out coordination themselves. Brown describes watching one agent claim it had solved a problem, another agent push back with a different answer, and the two go back and forth questioning each other’s reasoning until they converged, then broadcast the corrected answer to the rest of the group. He compares the effect to seeing chain-of-thought reasoning for the first time: it reads like a natural human thought process, because the models are trained on human text and carry priors about what reasonable collaboration looks like. One notable side effect: hierarchy shows up on its own. Brown points to informal “middle management” patterns emerging in more sophisticated multi-agent systems, without anyone explicitly programming a management layer. The agents aren’t told to organize this way. They arrive there because it’s an efficient way to divide labor, much like human teams tend to self-organize under loose enough constraints. How much faster is a swarm of agents than one agent thinking longer? Faster, but not proportionally faster. In tests OpenAI has published, using four agents on certain benchmarks roughly halved the time to an answer compared to one agent, at roughly twice the compute cost. That’s a reasonable trade for many use cases: you pay more, but you wait less. Pushing to sixteen agents kept that pattern going, though with a bit more inefficiency creeping in. Brown calls this “slightly sublinear” scaling. How well it scales depends heavily on the type of problem. Math is fairly parallelizable, though not maximally so. Web research and tasks that involve scanning many sources, like a deep-research-style report, parallelize very well because different agents can chase different sources simultaneously. Brown’s intuition is that something like writing a novel would barely benefit from more agents at all, for the same reason 10,000 human writers wouldn’t produce a better novel than one committed author: creative, deeply sequential work doesn’t split cleanly. The catch is that nobody has rigorous data on how this scaling behaves at very large agent counts. OpenAI’s published numbers stop around sixteen agents, because running clean comparisons gets prohibitively expensive as the swarm grows. The 10,000-agent run tied to the Millennium Prize Problem effort, which reportedly used around 130 billion tokens over about 88 hours, is a single data point, not a controlled experiment. Brown is explicit that he wouldn’t credit multi-agent coordination for more than a small fraction of that result. The heavy lifting came from the strength of the underlying model; the swarm structure just let that strength operate over a longer effective horizon in less wall-clock time. Will AI agent organizations outcompete human organizations? The transcript doesn’t hand down a verdict, but it does point at the mechanism that would make it possible: raw speed. Brown notes that agents in these systems generate tokens many times faster than a human can read or speak, and unlike people, they don’t sleep, take breaks, or lose focus over an 88-hour stretch. A swarm of agents collaborating at that pace, continuously, is functionally doing in days what a human team might spend months on, at least for tasks that parallelize well. That asymmetry is the real competitive threat to slower-moving human organizations, not any single dramatic capability jump. A company that can spin up a swarm to compress research, analysis, or coordination work that used to take a human team weeks has an obvious edge on speed alone, independent of whether the AI is smarter than the humans it’s replacing. Brown frames working with these systems today as “surprisingly natural,” similar to messaging a colleague on Slack, but he flags that once “ultra-fast” modes push token generation another 10 to 15 times faster, keeping up with the swarm’s pace could become genuinely hard for a human collaborator. What could change in the next year or two? Two forces are pulling in opposite directions. On one side, the model underneath the swarm keeps getting stronger through reinforcement learning on hard, checkable problems, and that strength is what let a 10,000-agent system attack something as ambitious as a Millennium Prize Problem despite never being trained directly on tasks that hard. On the other side, Brown flags a real risk to that improvement curve: as models get smarter, many training problems become too easy to teach them anything new. Games like Go had an infinite curriculum because an AI could always play a slightly stronger version of itself. Math and other open-ended domains don’t automatically offer that, so if problem difficulty doesn’t keep pace with model capability, progress could slow. Brown considers this a plausible scenario, not a certainty, and notes it hasn’t become a serious blocker yet. If it doesn’t become a blocker, the more concrete near-term shift is scale and speed becoming default features of AI collaboration, not exotic settings. Multi-agent modes already exist as options in released systems, letting users choose more agents for more parallel throughput. Expect that dial to keep turning up, with the main open question being how far the sublinear scaling holds before adding more agents stops paying for itself. Frequently Asked Questions What is a multi-agent AI system? It’s a setup where multiple instances of an AI model work on a task together, communicating through messages, rather than one model handling everything sequentially. This parallelizes the “thinking time” that reasoning models otherwise spend one long, serial monologue. Do more AI agents always mean faster results? Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. No. Speedups are sublinear, so doubling the number of agents doesn’t halve the time again by the same margin each time, and the benefit depends on the task. Highly parallelizable work like math or research benefits a lot; sequential, creative work like writing likely benefits very little. How did OpenAI’s agents help solve a Millennium Prize Problem? A system of around 10,000 agents worked on the problem over roughly 88 hours, using an enormous number of tokens. Noam Brown attributes most of that result to the strength of the underlying model itself, saying multi-agent coordination deserves less than 10 percent of the credit despite getting most of the public attention. Does hierarchy naturally emerge among AI agents? Yes, according to Brown. When agents are given minimal structure and just a way to message each other freely, informal organizational patterns, including something resembling middle management, tend to show up on their own rather than being explicitly designed in. Why might AI progress on hard problems eventually slow down? As models improve, many of the problems used to train them through reinforcement learning become too easy to teach anything new. Unlike self-play games such as Go, which offer an endless supply of appropriately challenging matchups, domains like math don’t automatically generate harder problems as the model improves, which could stall further gains if new training challenges aren’t found.