A.I.
Agent swarms
What thousands of AI agents accomplish together: speed, insight, and coordination, with every number called out.
Primer
A swarm is many agent instances pointed at one shared objective. They coordinate by messaging each other, they divide the labor, and they write shared artifacts the rest of the swarm can read.
That is different from a single agent with a harness. A harness is the computer, memory, and schedule around one model loop. Grok Bot and Muse are that story: one agent, or a small team, inside a harness. A swarm copies the loop many times and has the copies talk.
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.
Insight
Diverse attack angles, plus cross-checking. Separate agents can try different approaches, then test what the others return.
Coordination
Message passing, task decomposition, and handoffs from machine search to human-readable write-ups.
The stat row
Every swarm on this hub uses the same six fields. An unknown cell says "not disclosed." Where reports disagree, the cell stays a range. This page does not invent a figure, and it does not collapse a range into one number.
| Agents | Wall-clock hours | Tokens | Messages between agents | Est. compute cost | Human-equivalent hours |
|---|---|---|---|---|---|
| ~10,000 concurrent | 88 hours (Sept 1 to 5). Euler warm-up: ~100 agents, ~50 hours, before the scale-up. | 130B output tokens on the Navier-Stokes run (300B across the full effort) | 2.7 to 4.9M. Reports differ. | Not disclosed by OpenAI. Press estimates $2M to $22.5M. The Navier-Stokes run is commonly estimated at $6 to 10M. | not disclosed |
| Agents | Wall-clock hours | Tokens | Messages between agents | Est. compute cost | Human-equivalent hours |
|---|---|---|---|---|---|
| ~950 Claude agents (~949 sessions) | ~21 hours | 210M | not disclosed | not disclosed | Anthropic estimates weeks to months for an expert researcher by hand. |