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

This paper presents a domain-agnostic framework for resolving conflicts among multiple LLM-based agents over shared resources through bilateral negotiation, structured mediation, and procedural deconfliction, demonstrating its effectiveness in coordinating power grid applications for cost optimization and resilience.

The Gist

Imagine a busy kitchen where several expert chefs are working at the same time. One chef is trying to make the most expensive, high-end dish possible (let's call this the "Resilience Chef" who wants to save fuel for a storm). Another chef is trying to make the cheapest, fastest meal possible (the "Cost Chef" who wants to save money).

Both chefs need to use the same stove, the same oven, and the same limited supply of ingredients (the shared resources). If they don't talk to each other, they might both try to use the oven at the exact same time, or one might use up all the gas while the other is trying to cook. This is a conflict.

In the past, a strict head chef (a central computer) would have to step in, look at every single recipe, and force the chefs to stop doing what they wanted. But what if the chefs are actually AI robots (Large Language Models) that are very smart but also very private? They don't want to show their secret recipes (their internal logic) to the head chef, and they don't want to be micromanaged.

This paper introduces a new way for these AI chefs to solve their own arguments without a boss telling them what to do. They call it a "Deconfliction Framework."

Here is how it works, broken down into simple concepts:

1. The Problem: Too Many Smart Robots

Modern power grids (the electrical system that powers our homes) are getting crowded with smart apps. Some apps want to save money, others want to be ready for emergencies, and others want to use solar power. They all try to control the same batteries and generators.

• The Issue: If they all shout "Use the battery!" or "Don't use the battery!" at the same time, the system crashes or becomes inefficient.
• The Old Way: A central computer calculates the perfect solution. But this is hard because the apps are made by different companies, they speak different "languages," and they don't want to share their secrets.

2. The Solution: Giving Each App a "Negotiator"

Instead of forcing the apps to talk to a central computer, the authors give each app its own AI Agent (a digital negotiator).

• Think of these agents as diplomats.
• The "Cost App" has a diplomat who says, "We need to save money."
• The "Resilience App" has a diplomat who says, "We need to stay safe."
• These diplomats can talk to each other, reason through the problem, and find a middle ground without the apps having to reveal their secret formulas.

3. Three Ways to Settle the Argument

The paper tests three different "rules of engagement" for how these diplomats solve their fights:

A. The "Handshake" (Bilateral Negotiation)

• How it works: The two diplomats sit down face-to-face and talk directly. "I'll give you a little bit of battery if you give me a little bit of gas." They keep trading offers back and forth until they shake hands on a deal.
• The Vibe: Fast and friendly, but sometimes they might get stuck if they are too stubborn.
• Result in the paper: They found a solution very quickly (in 5 rounds of talking).

B. The "Referee" (Structured Mediation)

• How it works: The two diplomats don't talk directly. Instead, they both talk to a neutral Referee Agent. The Referee listens to both sides, calculates a fair compromise, and says, "Okay, here is a new plan that is 60% good for you and 60% good for you."
• The Vibe: More organized and fair, but it takes a little longer because the Referee has to do the math.
• Result in the paper: It took 9 rounds, but the result was very fair.

C. The "Traffic Light" (Procedural Deconfliction)

• How it works: There is no talking. There is just a strict, pre-programmed rulebook (like a traffic light). The system automatically calculates a "middle point" based on how flexible each side is being. If one side refuses to move, the system just picks the average.
• The Vibe: Very rigid and fast, but it doesn't allow for creative solutions.
• Result in the paper: Sometimes it failed to agree because the two sides were too far apart, and the rigid rules couldn't force them to meet in the middle.

4. The Real-World Test: The Power Grid

The authors tested this on a simulated power grid with diesel generators and batteries.

• The Conflict: The "Cost" agent wanted to drain the battery to save money on electricity prices. The "Resilience" agent wanted to charge the battery to be ready for a blackout.
• The Outcome:
  • The Direct Negotiation (Handshake) was the fastest. The agents realized that draining the battery a little bit was better than fighting to the death.
  • The Mediator (Referee) was the fairest. It made sure neither side got crushed.
  • The Rigid Rule (Traffic Light) sometimes got stuck because the agents just wouldn't budge.

Why This Matters

This is a big deal because the future of our power grid (and many other things like traffic control or logistics) will rely on many different AI systems working together.

• Privacy: The apps don't have to show their secret math to anyone.
• Autonomy: The apps get to keep making their own decisions, they just have to learn to compromise.
• Safety: It prevents the system from crashing when two smart apps try to do opposite things.

In a nutshell: This paper teaches us how to build a "diplomatic corps" for AI robots. Instead of having a boss force everyone to obey, we give the robots the tools to negotiate, compromise, and find a solution that works for everyone, keeping our lights on and our wallets happy.

Technical Summary

1. Problem Statement

Modern cyber-physical systems, particularly in electric power distribution, increasingly rely on semi-autonomous agents (e.g., Advanced Distribution Management Systems - ADMS, Distributed Energy Resource Management Systems - DERMS) that operate concurrently. These agents often propose conflicting control actions for shared resources (e.g., batteries, diesel generators) to optimize distinct, often opposing, objectives (e.g., cost minimization vs. grid resilience).

Traditional conflict resolution methods face significant limitations:

• Centralized Control: Assumes a single coordinator with global knowledge of all objectives, which is often unavailable due to privacy constraints or proprietary logic.
• Rigid Architectures: Require agents to be designed specifically to work together, limiting modularity.
• LLM Integration: The emergence of Large Language Model (LLM)-based agents introduces natural language reasoning and qualitative decision-making, but existing frameworks lack mechanisms to deconflict these agents without exposing their internal logic or forcing them to solve complex optimization problems directly.

The core challenge is to develop a domain-agnostic deconfliction framework that allows LLM-based agents to resolve conflicts over shared resources while preserving local autonomy, privacy, and the ability to use diverse decision-making strategies (quantitative and qualitative).

2. Methodology

The authors propose a framework where applications are wrapped by Client Agents (LLM-based) that negotiate over shared setpoints. The framework operates through three distinct Deconfliction Modes:

A. Agent Design Principles

Client agents are designed with specific constraints to ensure effective negotiation:

• Separation of Concerns: The agent is decoupled from the application's internal logic, interacting only via inputs/outputs.
• Privacy-Preserving: Agents negotiate without revealing their objective functions or private constraints.
• Chain-of-Thought (CoT): Agents use a structured reasoning process to evaluate proposals, justify decisions, and calculate compromises.

B. Three Deconfliction Modes

1. Bilateral Negotiation: Agents interact directly, exchanging offers and counter-offers iteratively until a mutual agreement is reached. This relies on direct communication and strategic compromise.
2. Structured Mediation: A dedicated Mediator Agent coordinates the process. It collects proposals, evaluates conflicts, and guides agents toward a consensus using an iterative weighted-consensus mechanism.
3. Procedural (Deterministic) Deconfliction: Replaces the mediator with a deterministic algorithm (adapted from game theory). It computes a weighted centroid of proposed setpoints.
   • Weighting Mechanism: Weights ($w_{i,k}$) are adjusted based on the flexibility demonstrated by each agent in previous iterations. Agents that move closer to the consensus centroid receive higher weights.
   • Convergence: The process iterates until convergence or a maximum iteration limit is reached. If no consensus is found, the final weighted centroid is imposed.

C. The Weighted-Consensus Mechanism

The framework utilizes an iterative update rule for setpoints ($x$) and weights ($w$):
$$c_k = \frac{\sum (w_{i,k} \cdot x_{i,k})}{\sum w_{i,k}}$$
Where $c_k$ is the centroid at iteration $k$. Weights are updated based on the Euclidean distance between an agent's initial proposal and the previous centroid, rewarding agents that demonstrate flexibility (moving toward the consensus) without requiring them to solve the global optimization problem explicitly.

3. Key Contributions

1. Domain-Agnostic Formulation: A framework treating applications as abstract agents capable of negotiating over shared resources, instantiated with three specific workflows (Bilateral, Mediated, Procedural).
2. LLM-Specific Design Principles: Definition of CoT-based reasoning structures for client agents that balance internal objectives with the incentive to compromise, ensuring privacy and modularity.
3. Iterative Weighted-Consensus: A novel mechanism that allows agents to reach consensus without solving global optimization problems, supporting both numeric and non-numeric (qualitative) decisions.
4. Empirical Evaluation: A comprehensive comparison of the three modes using a realistic power grid case study.

4. Experimental Results

The framework was tested on an IEEE 123-bus system with two conflicting applications:

• Cost Optimization: Minimizes operational costs (discharges batteries when grid prices are high, uses diesel).
• Resilience: Maximizes fuel reserves (charges batteries, minimizes diesel use).

Key Findings:

• Convergence Speed:
  • Bilateral Negotiation: Fastest convergence (5 rounds). Agents made strategic compromises early, reaching a solution with zero diesel generation and modest battery discharge.
  • Structured Mediation: Moderate convergence (9 rounds). The mediator dynamically adjusted weights based on cooperative behavior.
  • Procedural Deconfliction: Failed to converge within 10 rounds in specific trials. Agents with fundamentally incompatible goals (charging vs. discharging) refused to cross the threshold, leading to a stalemate where the final centroid was imposed.
• Performance Metrics (Success Score):
  • All three methods outperformed the "Initial Centroid" baseline (simple averaging).
  • Bilateral: High Pareto efficiency (solutions lie on the Pareto front) but exhibited high variance and lower fairness (unequal compromise).
  • Structured Mediator: Improved consistency and fairness (solutions closer to an equitable $x=y$ line) but slightly reduced Pareto efficiency compared to bilateral.
  • Procedural: Moderate performance, acting as a deterministic fallback.
• Outcome: All methods successfully guided agents toward a mutually beneficial operating region (minimal diesel, modest battery discharge) rather than a simple average of opposing extremes.

5. Significance

This work addresses a critical gap in the deployment of AI agents in critical infrastructure:

• Scalability & Modularity: It enables the integration of diverse, third-party applications without requiring them to be re-engineered for compatibility.
• Privacy Preservation: It allows agents to negotiate effectively without exposing proprietary algorithms or sensitive operational data.
• LLM Integration: It demonstrates how LLMs can be harnessed for complex, multi-objective decision-making in physical systems, moving beyond simple chat interfaces to active control and negotiation.
• Trade-off Management: The study highlights the inherent trade-off between Pareto efficiency (optimality) and consistency/fairness. While direct negotiation yields optimal results, it is volatile; structured mediation provides stability and fairness at a slight cost to optimality.

The paper concludes that while no single mode is perfect, the framework provides a robust, flexible foundation for managing autonomous agents in shared-resource environments, with future work focusing on tuning structural parameters to improve Pareto efficiency in mediated scenarios.