In this work, the authors construct a persistent multi‑disciplinary team of AI researchers, each maintaining a private research log and sharing experimental results, while an adaptive allocation mechanism assigns research opportunities based on the evolving evidence. The team is tasked with autonomously discovering linear attention architectures and is evaluated on a 1.3 billion‑parameter language model benchmark that includes common‑sense reasoning and both synthetic and real‑world retrieval tasks. The resulting architecture surpasses the GDN2 baseline across all metrics, demonstrating superior performance in language modeling, reasoning, and retrieval. Controlled ablation studies show that removing any single design element degrades outcomes: teams lacking disciplinary diversity plateau early, teams without private histories make slower progress despite identical shared evidence, and teams using uniform researcher selection under the same experimental budget yield fewer discoveries than those employing adaptive allocation. These findings indicate that autonomous scientific discovery hinges not only on agent capability but also on the structure of the research team, the diversity of its members’ expertise, the preservation of individual research histories, and the strategic allocation of limited experimental resources.