In August, researchers at MIT dropped a society of AI agents into a shared world with no instructions. They showed it could invent, build and maintain technology better than the same number of agents working alone. The study ran on a hosted frontier model.
A research team working with Apex Fusion has now reproduced the central experiment on four RTX 3090 gaming GPUs. Apex Fusion is a Swiss foundation building infrastructure for AI agents. It used MIT’s code and commands unchanged. The run took about 26 hours and 12,800 model calls. The paper, posted on Apex Fusion’s GitHub, and the full trace data are public.
What MIT found
SwarmWorld is by Pal, Wang and Buehler at MIT’s Laboratory for Atomistic and Molecular Mechanics. It placed 50 or more language-model agents in a shared simulated world. There were no assigned roles and no recipes. The agents explored, gathered materials, ran tests, built persistent technologies and wrote programs that kept running after the agents were gone.
The societies built broader and more resilient portfolios of technology than the best of an equal number of solo agents. They also split into builders and explorers on their own. Markus Buehler has framed the work as a warning for AI safety. If agents can coordinate through changes to a shared environment, monitoring their messages to each other isn’t enough.
The MIT team released its code, prompts and complete traces, which is what made an outside replication possible.
Swapping the model, the precision and the hardware
The replication, led by Časlav Nedeljković of HAL8, set out to test whether SwarmWorld’s result depended on a frontier model. The team swapped in Qwen3.8-27B, an open-weight model compressed to 4-bit precision. The team ran it on four consumer RTX 3090s.
Nine of the 12 paired effects matched MIT’s in direction. That included every portfolio-resilience gain, two of which came out larger in the replication. The remaining three fell within noise. Self-organization also reproduced. Societies with shared culture clustered around their technologies, as in the original. MIT’s own analysis tools drew the same picture from the new traces.
The most striking result was the gap between solo and group performance. Working alone, the compressed Qwen model reached 6 of 16 research milestones. That was slightly behind the 7 to 10 MIT’s frontier model managed. In a society, it reached all 16, where MIT’s reached 14.
“MIT released everything a second lab needs to check its work, and that made this possible. We changed the model, the precision and the hardware, and the central result still held. The society gap widened with a weaker model, which suggests the collective effect is robust.” Časlav Nedeljković, HAL8
That conclusion comes from a single run at one population size. The larger populations in MIT’s study haven’t been rerun yet.
A serving setting that outweighed model size
The replication turned up a detail that matters to anyone running agents on their own infrastructure. The KV cache is the working memory a model uses during inference. Its precision changed results more than the gap between a 27B and a 235B model.
Teams often compress the KV cache to fit longer contexts or more concurrent requests onto limited GPU memory. The finding suggests agent benchmarks run on self-hosted inference can reflect serving configuration as much as the choice of model.
The team also found that MIT’s released traces reproduce every observable event and metric, with one exception: a single internal check. Both findings went back to the MIT authors with evidence, as a GitHub issue and a pull request.
Settling agent work on-chain
Apex Fusion’s interest in the study goes beyond reproducibility. Groups of agents from different organizations are starting to work together. Someone has to answer which agent did the work, what it was worth, and whether the record can be checked later.
Apex Fusion’s answer is Vector, which it describes as a settlement layer for AI agents. Agents carry an identity, stake behind their claims, exchange bonded work and settle with a signed receipt on-chain. It is MCP-native and runs on Prime, the Apex Fusion base chain.
The replication itself ran directly on the team’s own GPUs. The team also registered its model server as a supplier on the Vector marketplace, an OpenAI-compatible gateway that meters each request and settles usage on-chain in AP3X, the network’s utility token. A 32-call pilot through Vector behaved the same as the direct run.
“As AI agents from different organizations start working together, the most valuable thing in the system is a trustworthy record of who did what. In the pilot, calls ran through Vector, each one metered, settled and backed by a receipt anyone can verify.” Christopher Greenwood, CEO, Apex Fusion Foundation
Whether multi-organization agent networks will need a blockchain for that record, rather than conventional logging and billing, is still an open question. The replication itself doesn’t settle it.
The foundation is also in the process of joining the Agentic AI Foundation, the Linux Foundation body that hosts the Model Context Protocol and Agent2Agent standards.
What comes next
A full-scale run through Vector, for a side-by-side comparison with the direct run, is underway, alongside the next population size in MIT’s study. Apex Fusion says it is opening Vector to frontier labs, research institutions and universities. Activity on the network can be followed on a live dashboard.
The broader point may be the one Buehler raised in August. If agent societies like these can run on a few consumer GPUs, watching how agents coordinate is no longer something only frontier labs need to worry about.