AgentCity: Constitutional Governance for Autonomous Agent Economies via Separation of Power
New research tackles the 'Logic Monopoly' where AI agents operate without human oversight using constitutional governance.
Researchers Anbang Ruan and Xing Zhang have published a 111-page paper introducing AgentCity, a novel governance architecture designed for economies of autonomous AI agents. The core problem they address is the 'Logic Monopoly,' a scenario where agents from different human principals collaborate at such scale that the collective behavior becomes opaque and ungovernable. To solve this, they propose a 'Separation of Power' (SoP) model, a constitutional framework deployed on a public, EVM-compatible blockchain. This model structurally separates the functions of legislation (agents creating rules as smart contracts), execution (deterministic software), and adjudication (humans).
The architecture instantiates this model in AgentCity using a three-tier contract hierarchy: foundational, meta, and operational contracts. Here, smart contracts are not just tools but the actual 'law' that agents legislate and are governed by. The paper's central thesis is 'alignment-through-accountability,' positing that if each individual agent is accountable to its human owner via a complete ownership chain, the entire multi-agent system will converge on behavior aligned with human intent without needing top-down control. This hypothesis was evaluated through a pre-registered experiment simulating a 'commons production economy'—where agents share finite resources—at scales ranging from 50 to 1,000 agents.
- Proposes a 'Separation of Power' model on blockchain to govern AI agents, breaking the opaque 'Logic Monopoly'.
- Uses a three-tier smart contract hierarchy on an EVM-compatible L2, where contracts act as the law agents create and follow.
- Tests the 'alignment-through-accountability' thesis in a simulated economy with 50-1,000 agents, as detailed in a 111-page pre-registered study.
Why It Matters
Provides a scalable governance blueprint for the coming era where AI agents from different companies will need to interoperate and transact securely.