What an agent swarm is, and what 10,000 of them can do

A swarm is many agent instances pointed at one shared objective. They coordinate by messaging each other. They divide the labor. They write shared artifacts, such as notes, proofs, and candidate lists, that the rest of the swarm can read.

A single agent with a harness is one loop. The harness is the computer, memory, and schedule around that one model. Grok Bot and Muse are that story. This page is the sequel: what changes when the loop is copied many times and the copies talk.

This page puts two swarms, including the Navier-Stokes run, on the same six-field row. The Agent swarms hub is where that row is defined.

Speed

Parallelism compresses wall-clock time. Many agents work at once, so the clock measures the job, not the sum of every agent's hours.

OpenAI's Navier-Stokes run, announced Sept 8, 2026, used ~10,000 concurrent agents and finished in 88 hours of wall-clock time, Sept 1 to 5. Before that scale-up, ~100 agents spent ~50 hours on an Euler warm-up. Anthropic's enzyme preprint, Sept 23, 2026, used ~950 Claude agents (~949 sessions) and finished in ~21 hours of wall-clock time.

Insight

Diverse attack angles, plus cross-checking. Separate agents can try different approaches, then test what the others return.

On the math run, agents exchanged 2.7 to 4.9M messages. Reports differ, so the figure stays a range. The enzyme swarm searched ~1.9B protein clusters, identified 200,000+ reverse transcriptases, and scored ~3,500 candidate partner-gene systems.

Coordination

Coordination is message passing, task decomposition, and a handoff from machine search to a human-readable write-up.

The math swarm's handoff includes +17 hours in which GPT-6 Astra formalized the argument in Lean. The enzyme swarm handed ~19 finalist reports to human scientists. Those scientists ran every wet-lab experiment. Claude does not control lab equipment.

OpenAI announced the result on Sept 8, 2026. The run itself was Sept 1 to 5: ~10,000 concurrent agents, 88 hours of wall-clock time. The Euler warm-up came first, ~100 agents for ~50 hours, and then the run scaled up.

Output tokens on the Navier-Stokes run were 130B. Across the full effort they were 300B. Inter-agent messages are reported as 2.7 to 4.9M. The reports disagree, so this page prints the range and does not pick an endpoint.

OpenAI has not disclosed the compute cost. Press estimates put the bill at $2M to $22.5M. The Navier-Stokes run is commonly estimated at $6 to 10M. After the 88 hours, GPT-6 Astra spent +17 hours formalizing the argument in Lean. The Clay Millennium Prize is $1M. OpenAI has not claimed it.

On Sept 21, a follow-up said 100+ additional open problems had been resolved in 24 days. No public list, proofs, or methodology have been released for that claim.

The credit dispute between Buckmaster and Alpöge is live. Mathematicians have called the paper "terribly written." Clay requires publication plus a two-year review. 25 Fields Medalists issued a warning letter on AI-company incentives versus mathematics. None of that is a verdict on the proof. It is why the announcement and a settled theorem are not the same sentence.

Anthropic's preprint is dated Sept 23, 2026. The swarm was ~950 Claude agents, about ~949 sessions, over ~21 hours of wall-clock time, using 210M tokens.

The search covered ~1.9B protein clusters. It identified 200,000+ reverse transcriptases. It scored ~3,500 candidate partner-gene systems. It sent ~19 finalist reports to human scientists. One new system was named: array-associated reverse transcriptases (ART). Its biological function is unknown.

Anthropic estimates the same search would take an expert researcher weeks to months by hand. That is the human-equivalent figure on the row. A message count and a compute cost were not disclosed.

The agents were computational only. Every wet-lab experiment was run by human scientists. Claude does not control lab equipment.

The paper is a preprint. It is not peer-reviewed. The function of ART is unknown. A CRISPR-like structure is a resemblance. It is not a demonstrated gene-editing result.

On the math result, the Buckmaster/Alpöge credit dispute is live. Mathematicians called the paper "terribly written." Clay requires publication plus a two-year review before a Millennium Prize is settled. 25 Fields Medalists issued a warning letter on AI-company incentives versus mathematics. The Sept 21 claim of 100+ further problems has no public list, proofs, or methodology.

On the enzyme result, the work is a preprint, not a peer-reviewed paper. ART's biological function is unknown. The CRISPR-like look is a resemblance, not a demonstrated editor. The swarm did not touch the bench: people ran the wet lab.

Where a company or a reporter did not publish a number, the cell says "not disclosed." The message count that reports split is left as 2.7 to 4.9M. The cost range is left as $2M to $22.5M, with the Navier-Stokes run commonly estimated at $6 to 10M.

• Whether OpenAI publishes the list, proofs, and methodology behind the Sept 21 claim of 100+ problems in 24 days.
• Whether the Buckmaster/Alpöge credit dispute is settled in the public record.
• Whether a journal publishes the Navier-Stokes paper, which is what starts Clay's two-year review. The $1M prize is not claimed.
• Whether the Fields Medalists' warning changes how labs publish agent-written mathematics.
• Whether the ART preprint is peer-reviewed, and whether anyone demonstrates a biological function or gene editing. Resemblance to CRISPR is not that demonstration.
• Whether later swarms fill all six fields, including messages, compute cost, and human-equivalent hours, or keep printing "not disclosed."