Princeton paper: combined specialist agents from many developers beat general agents on GAIA

A Princeton paper reports that combining specialized AI agents built by different developers reached 70% accuracy on the GAIA benchmark, while each specialist alone scored at most 35%.

The paper says that under the same total development budget, the combined system beat the median general-purpose agent by 12 points.

On SealQA, a benchmark limited to web search, the combined system performed worse than the general agent and cost more, according to the same report. The source for these figures is a post by Rohan Paul summarizing the paper, not the paper itself.