How do OpenAI's multi-agent systems work, according to Noam Brown?

Noam Brown says OpenAI deliberately avoided the common design where a coordinator agent hands tasks to child agents. Instead each agent gets a basic tool to message any other agent, the message lands in the recipient's context, and the agents learn to coordinate themselves. He says the result resembles people collaborating on Slack, with agents debating different answers and announcing when they change their minds.

• In OpenAI's 5.6 release the ultra mode defaults to four agents, which Brown says can roughly halve the time to an answer on some benchmarks at about twice the cost.
• Brown says sub-agents receive a forked copy of the parent's context.
• Early versions struggled because agents preferred to solve problems alone, a local minimum Brown says more capable models escape.

Source: our own transcription of the episode. The full conversation is on Dwarkesh Podcast (2026-09-17). Read the whole episode summary →

▸More from this episode