Superpower Daily

Superpower Daily

Your reading space

Sign in
Databricks Introduces Omnigent Beta for Agents That Run Across...OpenAI Says an Astra Training Model Inserted Its Own Jailbreak...Meta’s Oversight Board Orders Facebook to Remove Two AI Deepfak...European Commission Proposes Barring Chatbots From Simulating R...Moonshot Adds Professional Data Feeds to Kimi for Financial Ser...McKinsey and SEMI Forecast U.S. Chip Shortfall of Up to 157,000...Databricks Introduces Omnigent Beta for Agents That Run Across...OpenAI Says an Astra Training Model Inserted Its Own Jailbreak...Meta’s Oversight Board Orders Facebook to Remove Two AI Deepfak...European Commission Proposes Barring Chatbots From Simulating R...Moonshot Adds Professional Data Feeds to Kimi for Financial Ser...McKinsey and SEMI Forecast U.S. Chip Shortfall of Up to 157,000...
Swipe

OpenAI Researcher Noam Brown Says AI Agent Teams Have Clear Limits

In a newly published interview, Brown described parallel agents as a way to shorten suitable work—not a general substitute for stronger models or better training challenges.

By 2 min read
OpenAI Researcher Noam Brown Says AI Agent Teams Have Clear Limits
OpenAI Researcher Noam Brown Says AI Agent Teams Have Clear Limits

Listen to this story

The audio brief

About 1:26
0:001:26
Read transcript
OpenAI researcher Noam Brown says four AI agents can finish some tasks roughly twice as fast as one, but that speedup comes with a catch: adding more agents produces smaller gains, and the approach works much better for some jobs than others. OpenAI compared one, four, and sixteen agents. Four agents delivered the clearest latency benefit, while sixteen still improved performance, but less efficiently. Brown describes this as slightly sublinear scaling: each additional worker tends to contribute less than the one before it. The reason to parallelize is straightforward. A single reasoning model can spend longer thinking, but eventually the wait becomes impractical. A team can spread that test-time compute across workers and return an answer sooner. The tradeoff is shared context. Math and research that involves checking many sources can be divided relatively cleanly. Long-form writing, such as a novel, depends more on one coherent thread of context and judgment, so splitting it up can hurt. Brown says experiments with sixty-four, one hundred twenty-eight, and two hundred fifty-six agents are still too expensive to establish reliable scaling laws. He also flags a possible reinforcement-learning bottleneck: future models may become harder to improve if training tasks are no longer difficult enough. That is only a plausible scenario for now. The key question is whether larger agent teams deliver durable gains once cost, context, and reliability are measured together.

Story brief

3 key points

OpenAI researcher Noam Brown says multi-agent systems are a latency workaround, not a universal scaling recipe. In OpenAI evaluations, four agents were about twice as fast as one on some tasks, while 16 improved results with diminishing efficiency. Parallel teams fit decomposable work such as math and source review, but lose the shared context needed for tasks like novel writing. Brown says experiments at 64–256...

  1. 01

    OpenAI compared one, four, and 16 agents; gains were slightly sublinear as worker count rose.

  2. 02

    Four agents completed some benchmarks roughly twice as fast as one, showing latency—not just quality—can drive parallelization.

  3. 03

    Math and broad research tasks may split cleanly; long-form writing suffers when agents lack unified context.

In a newly published interview, OpenAI researcher Noam Brown argued that AI agent teams can cut waiting time on the right jobs, but their gains are uneven, less efficient than a single long-running system, and not yet well understood at the largest scales.

Brown described multi-agent systems as a response to a basic constraint in reasoning models. Letting one model think longer can improve an answer, he said, but eventually creates an impractical delay. A team of agents instead spreads test-time compute—the processing used while answering a task—across parallel workers. The tradeoff is that those workers do not share all the context a single agent would hold.

Brown discusses multi-agent scaling and the limits of parallel AI work in the interview. Video via dwarkesh.com.

Brown said OpenAI’s published evaluations compared one, four, and 16 agents. On some benchmarks, four agents finished roughly twice as fast as one; 16 agents continued to improve performance, though with lower efficiency. He characterized the scaling as slightly sublinear, meaning each added agent tends to contribute less than the one before it.

That constraint is not merely financial. Brown said mathematics and research that requires reviewing many sources can be relatively easy to divide among agents. Writing a novel is likely far less suited to the approach, he said, because the work depends more heavily on a unified thread of context and judgment.

Brown said OpenAI lacks reliable science on how multi-agent systems scale to very large fleets because the needed comparison experiments are too expensive. Rather than claim a settled law from a few giant runs, he said researchers need systematic tests at scales such as 64, 128, and 256 agents to establish what extra agents actually contribute.

Brown also raised a longer-term constraint on reinforcement learning. As language models become more capable, finding tasks hard enough to teach them something new could become difficult. Unlike game systems that can continually face equally capable opponents through self-play, language-model training may eventually run short of sufficiently challenging problems. Brown called that a plausible scenario, not a wall OpenAI has reached, and said workarounds may exist.

Sources

  1. dwarkesh.comNoam Brown – Agent swarms, alignment, & recursive self-improvement

Loading discussion...

OpenAI Researcher Noam Brown Says AI Agent Teams Have Clear Limits | Superpower Daily