To overcome limitations of individual generative AI models, recent multi- agent systems methods have been introduced for how generative AI agents can invoke external functions. We analyze the relationship between these paradigms and existing methods for developing distributed cognitive architectures as illustrated by the DIARC cognitive robotic architecture. We show how exposing DIARC actions as "MCP tools" and using DIARC's TRADE middleware as a substrate for "A2A-style communication" enables integration of cognitive robotic architectures into agentic frameworks, allowing AI agents to utilize existing cognitive functions and architectures while providing cognitive architectures with generative AI modules in a systematic way. We further argue that semantic action representations and reflective middleware monitoring provide key ingredients for safer and more introspectable multi-agent systems, especially in mixed-initiative and embodied domains, because they allow cognitive systems to check, mediate, and supervise the interactions among agentic components rather than merely relaying them.
@inproceedings{scheutzkrause26aaaiss,
title={Bridging Cognitive Robotics and Generative AI Protocols},
author={Matthias Scheutz and Evan Krause},
year={2026},
booktitle={Proceedings of AAAI Spring Symposium on Multi-Agent System Safety and Teamwork (MASST)},
url={https://hrilab.tufts.edu/publications/scheutzkrause26aaaiss.pdf}
}