Research shows two AI agents can match complex transformer models

Researcher Sergey Salishev has published a groundbreaking paper titled *Transcript-Managed Transformers: Monotone Multi-Agent Collapse and Universality with Two Pop-Enabled Transcripts* on arXiv, exploring how transcript management in fixed-precision causal Transformers can simplify complex AI behavior.

Salishev introduces the concept of transcript-managed transducers, showing that by using two pop-enabled transcript channels, it's possible to achieve universality in Transformer models. The work demonstrates that a monotone multi-agent protocol—where agents append, route, and copy visible blocks—can collapse to deterministic finite-state transducers. For $k \geq 2$ pop-enabled channels, the system achieves Turing completeness (RE), while a single channel ($k=1$) matches deterministic context-free languages (DCFL). This suggests that multi-agent AI systems can be simplified without losing computational power.

The paper also highlights that orchestrated one-channel agents can match a single controller with $k$ channels, implying that two pop-enabled transcripts—whether in one agent or two—suffice for universality. The findings challenge traditional assumptions about Transformer complexity and open new avenues for designing more efficient multi-agent AI systems.