Why We Need This

The Governance Challenge for AI Agents

Artificial Intelligence (AI) Agents, particularly those powered by Large Language Models (LLMs), hold immense potential for participating in and even automating aspects of governance, especially in decentralized systems. However, deploying them reliably in such contexts faces significant hurdles:

Consensus

LLMs are inherently probabilistic. For a given input sequence, an LLM calculates probability scores for potential next words, leading to a vast number of possible output sequences. While mechanisms like greedy search or beam search select a likely output (the red path in the diagram below), this selection isn't inherently deterministic or universally agreed upon. In governance scenarios, especially those involving asset allocation or critical decisions, how do we ensure that one specific output path is the authoritative and accepted result, agreed upon by all stakeholders? Relying solely on trusted execution environments (TEEs) isn't sufficient; the process of generating the AI's decision requires its own form of verifiable consensus.

Image illustrating multiple potential output sequences from an LLM

Identity(Consistency)

Even if consensus on an output is achieved, distributed systems face concurrency challenges. Multiple valid, yet potentially contradictory, decisions could theoretically be generated simultaneously for the same Agent regarding the same issue. This is analogous to the double-spending problem in blockchains, which is resolved by enforcing a single, linear chain of transactions. An AI Agent involved in governance cannot simultaneously provide contradictory directives, just as a human leader cannot credibly issue opposing orders at the same time. The Agent must exhibit a singular, consistent identity and state at any given point in time.

Continuity

Human societies rely on memory for continuity. Decisions, commitments, and identities persist over time. An AI Agent participating in governance must also possess a reliable form of memory. Its past interactions, decisions, and learned state must inform its future actions consistently. An Agent that "forgets" its previous commitments or rationale loses its credibility and effectiveness, much like an elected official forgetting their campaign promises.

Consensus, Identity, and Continuity are therefore fundamental properties required for AI Agents to function reliably and trustworthily within governance frameworks, particularly in distributed environments.