Collaborative LLM-based Agents for Modeling and Simulation

Modeling and simulation (M&S) offers a wide range of benefits to the U.S. Air Force, impacting everything from training and mission planning to developing new technologies. A significant challenge with M&S is developing intelligent and realistic agents that can simulate blue teammates and red adversary threats in these simulations. Making it easier and less expensive to create safe, secure, and reliable agents that work well together would significantly enhance the utility of existing M&S investments.In response to this need, Stottler Henke proposes to develop theĀCollaborative LLM-based Agents (COLA)Āsystem, an agentic workflow that quickly creates safe, secure, and reliable teams of LLM-based software agents. Given a set of instructions along with a knowledge base (doctrine, policy, instructions, training materials, commanderÆs intent) and an interface that supports observations and actions relevant to the simulated role, COLA will direct an LLM-based agent to fill the role of a human within a M&S environment. Planned Phase I research begins with literature review and verifying requirements for the proposed workflow. Based on the results of this work, we will create a detailed system design, develop a limited proof-of-concept prototype, and demonstrate key aspects of COLA within a relevant simulation.