Agentic Workflows for Resolving Conflict Over Shared Resources: A Power Grid Application

A research team led by Shiva Poudel has introduced a formal framework for deconflicting the actions of multiple LLM-based agents operating on shared resources. As AI agents are increasingly deployed for autonomous decision-making in domains like infrastructure control, their independent goals can lead to direct conflicts—such as two systems trying to command the same power generator at once. This new framework provides a systematic way to resolve these clashes through three escalating modes: starting with direct agent-to-agent negotiation, moving to a structured mediation process, and finally applying a deterministic procedural rule if consensus isn't reached. A key innovation is an iterative weighted-consensus mechanism that finds solutions without requiring each individual application to solve complex optimization problems, making it more efficient and scalable.

The framework's effectiveness was demonstrated in a critical real-world scenario: managing a power distribution grid. Researchers tested it with conflicting Advanced Distribution Management System (ADMS) applications—one focused on cost optimization and another on grid resilience—both vying for control of shared resources like diesel generators and Battery Energy Storage Systems (BESS). The system successfully coordinated these agents, ensuring reliable and cost-effective operation. The authors emphasize that the design is domain-agnostic, supporting both numeric and non-numeric decisions, which opens the door for its application beyond energy to areas like traffic management, cloud computing resource allocation, and multi-agent robotics.