Resilient Consensus in Agentic AI
Sribalaji C. Anand, George J. Pappas
Why It Matters
What makes this one worth your time
Understanding the consensus capabilities of LLM agents is crucial for their safe deployment in multi-agent systems, especially in adversarial contexts.
This research reveals the limitations of LLM agents in achieving consensus and suggests classical filtering methods as a potential solution.
Summary
The paper investigates the application of classical resilient consensus theory to large language model (LLM) agents in multi-agent systems, demonstrating that these agents can fail to reach consensus even in theoretically favorable conditions, while also showing that classical filtering methods can enhance agreement.
Key contributions
- Empirical evaluation of LLM agents in the context of Byzantine consensus.
- Demonstration of the limitations of LLM agents in reaching consensus under controlled conditions.
- Analysis of the impact of classical filtering methods on LLM agent agreement.
Notable insights
- LLM agents can fail to achieve consensus even when classical theory suggests it is possible, indicating unique challenges in their deployment.
- The effectiveness of classical resilient consensus filters varies based on the robustness of the underlying communication topology.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2606.15024v1 Announce Type: cross Abstract: Large language model (LLM) agents are increasingly deployed in multi-agent systems where they must coordinate and agree on shared decisions. We ask whether classical resilient consensus theory, developed for deterministic agents, transfers to LLM agents that may behave adversarially. Framing LLM agreement as a Byzantine consensus game, we run controlled experiments on complete and general communication graphs. We find that prompted LLM agents fail to reach agreement that is achievable in principle: consensus can fail even in settings where classical theory guarantees that a convergent algorithm exists, and this failure persists across temperatures and horizons. At the same time, wrapping the agents with classical resilient consensus filters improves agreement. The benefit of filtering depends on how much robustness the underlying topology already provides. Our results suggest that classical resilient consensus theory is a useful lens for the safety of agentic AI.