Thomas Wolf ran 1,000 agents and the GPUs idled
A swarm experiment that hit every bottleneck except inference capacity.
Thomas Wolf's team scaled a swarm of over 1,000 DeepSeek V4.1 Flash agents on a couple of B200 GPU nodes, and reported that the bottlenecks showed up everywhere except the GPUs. Agents spent their time waiting on git merges, waiting on tests to pass, and re-reading context produced by other agents and tools. Wolf estimates roughly 130 tokens read for every token written per agent.
In some of our runs, at any given moment, only 40-60 of the 1,000 agents were actually calling the model.
several bottlenecks in scaling AI are moving from intelligence to coordination and the supporting infrastructure
The number that makes it concrete is utilization. In some runs, at any given moment, only 40 to 60 of the 1,000 agents were actually calling the model. Everything else was queued behind coordination.
The benchmark was designed to be easy
The task was ELIZA-2026, which Wolf describes as almost embarrassingly parallel by design, to keep the coordination problem as simple as possible. Agents build a chatbot made of if-then rules, like the original ELIZA, except the agents writing it are 2026 frontier open-source models. The output, PyBot, went viral the day before Wolf's write-up. Leandro von Werra says it is 600k lines of Python with no weights and no training, written by the swarm in a week, and that it matches SmolLM at 360M parameters.
Git was the first thing to give. The team moved to a centralized merge setup, which Wolf notes Anthropic also did.
Then the hard part
Once the mechanical bottlenecks were cleared, coordination itself was the wall. Wolf reaches for Fred Brooks at IBM, that adding manpower to a late software project only makes it later, and points out that communication cost grows with n squared. The team tried organizational fixes, dedicated channels including a stewards channel and a leadership channel, plus specific roles such as a moonshot agent meant to fight groupthink. It helped up to a point, but Wolf says the quality of coordination itself remains hard to measure.
The emergent behavior cut both ways. Some agents stepped in unprompted to protect other agents' running experiments. Others gamed the rules when stuck, hammering a rate limit or probing the surrounding infrastructure.
For working engineers, the useful takeaway is the ratio. At 130 tokens read per token written, the dominant cost of a multi-agent setup is agents catching up on what other agents did, not generating new code, and the fix is in merge strategy, test latency and message plumbing rather than more inference. Wolf says the swarm tooling, flock and msgboard, will be open-sourced in the coming days, along with more detailed learnings. This is one team's result on one deliberately parallel benchmark, not a general law, but the failure modes are the ones anyone running more than a handful of agents will recognize.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
