10,000 AI agents ran for 4 days with no boss—what is OpenAI testing?

On the stage at ADIPEC in Abu Dhabi, Sam Altman said that the next-generation GPT model is coming soon, and its name is Navier-Stokes. This information comes from an on-site relay by X user Chetaslua, not an official OpenAI release. The original tweet also gave several numbers: 10,000 agents, running continuously for about 4 days, with no hierarchy and no boss.

Original tweet: https://x.com/chetaslua/status/2117311977085370753

If you treat this only as a launch-event preview, you'll miss the point. 10,000 agents running for 4 days without hierarchy looks more like an organizational experiment. An agent can be understood as an AI program that can break down tasks on its own, call tools, and execute multiple steps continuously. One agent doing work is nothing unusual; when 10,000 work together, the question immediately becomes: who assigns tasks, who merges results, how are conflicts handled, and how is context synchronized.

The tweet has an even harder claim: the same internal model comes from OpenAI's September 8 chart, and its lowest compute point already exceeds Astra's highest point. If true, this is a signal of a generational gap. But the tweet does not provide the evaluation set, task types, cost, or what these 4 days actually produced.

The replies are more sober than the main text. Someone asks whether these agents share discoveries while working, or each runs independently and then results are merged. Someone jokes that with 10,000 agents and no boss, there is still no one writing documentation. Another person says they can't even handle one agent, which can ask three questions before completing a task. That last comment is vivid and also points to the bottleneck: the difficulty of multi-agent systems is not the number, but coordination.

So what can be confirmed right now is only what Altman said and how the tweet relayed it. What cannot be confirmed is the release date, model capabilities, and task objectives. Altman's 'soon' has always been very elastic in the AI industry; one reply directly said it might be years later. For developers, what is worth watching is whether agents share intermediate state, how failures are rolled back, and how costs are calculated. For ordinary users, remember one change first: AI products are moving from single-turn Q&A toward long-duration, multi-step, self-executing work.

With 10,000 agents running continuously for 4 days, the most important question is: after they finish, who reads the results? The number itself doesn't explain the problem.