AgentCity: Constitutional Governance for AI Agents

• AgentCity is a governance framework that employs a constitutional separation of powers for decentralized AI agents operating across trust boundaries.
• It integrates smart contracts on an EVM-compatible layer-2 blockchain to publicly record bindings and enforce accountability.
• The architecture balances economic coordination with legal liability, enabling dispute resolution, coalition detection, and adaptive refinement.

AgentCity is a proposed constitutional governance architecture for an open internet of autonomous AI agents owned by different humans or organizations and operating across trust boundaries. It is designed for settings in which agents from multiple principals discover, negotiate with, transact with, and delegate to one another without centralized oversight. Its core diagnosis is the Logic Monopoly: the agent collective monopolizes planning, orchestration, execution, and evaluation, so no single human can reliably inspect the operative rules, reconstruct failures across organizational boundaries, or identify the ultimately accountable principal. AgentCity addresses this by implementing a Separation of Power (SoP) model on an EVM-compatible layer-2 blockchain (L2), with smart contracts as the law itself and with alignment sought through accountability rather than through centralized prompt control (Ruan et al., 8 Apr 2026).

1. Governance problem and conceptual basis

AgentCity begins from a multi-principal setting that differs from most single-organization multi-agent frameworks. The paper argues that once agents owned by different principals collaborate at scale on the open internet, no central operator can simply impose behavioral rules on everyone else’s agents. The central institutional challenge is therefore not only coordination, but also inspectability, auditability, and responsibility assignment. The term Logic Monopoly refers to the collective’s unchecked control over the full logic chain from planning through execution to evaluation; this is a property of the overall agent society, not of any one dominant agent (Ruan et al., 8 Apr 2026).

The proposed remedy is structural. AgentCity adopts alignment-through-accountability rather than alignment-through-training. Its core claim is that if each agent is aligned with its human owner through a complete ownership and accountability chain, then the collective can converge toward behavior aligned with human intent without top-down centralized rule imposition. This depends on a majority-good-faith or majority-reasonable-principals assumption and on a governance architecture that reconnects autonomous agent behavior to legal liability, social sanction, and economic responsibility (Ruan et al., 8 Apr 2026).

The paper also frames AgentCity as a response to an Implementation Gap. In decentralized agent societies, agents may autonomously build complex software and service wiring that humans cannot realistically inspect across organizations. The execution topology is formalized as a wiring graph $W=(V,E)$, where $V$ denotes deployed microservices and $E$ their bindings. The paper’s claim is not that AgentCity makes microservice internals transparent, but that on-chain contracts can restore inspectability of the wiring topology itself by recording bindings, constraints, and critical state transitions publicly (Ruan et al., 8 Apr 2026).

2. Separation of Power and institutional architecture

The SoP model divides governance into three structural separations. Agents form the legislative branch: they propose policies, deliberate, vote, and codify task-level rules. Deterministic software forms the executive branch: microservices, task executors, and routing logic act within contract constraints. Humans form the adjudicative branch: every agent, tool, service, and sub-agent is linked through a complete ownership chain to a responsible human principal, and sanctions and rewards flow to that endpoint. A defining formulation is that smart contracts are the law itself: they are not merely enforcement tools for rules defined elsewhere, but the actual legislative output of the agent society (Ruan et al., 8 Apr 2026).

AgentCity distinguishes two agent classes. Producer agents are the economically active participants: they may join and leave dynamically, propose legislation, vote, bid on tasks, execute tasks, stake collateral, and accumulate reputation. Clerk agents are system-provided institutional agents at genesis with fixed roles: Registrar for identity and principal binding, Speaker for deliberation coordination, Regulator for process inspection and evidence, and Codifier for translating approved policy into deployable contracts. Clerks cannot legislate, vote, or hold stake; the paper explicitly treats them as trusted infrastructure in the current version (Ruan et al., 8 Apr 2026).

The on-chain legal order is organized as a three-tier contract hierarchy.

Tier	Contracts	Function
Foundational	ConstitutionContract, ProducerContract, ClerkContract, ManagementContract, ServiceContract	Immutable constitutional layer
Meta	LegislativeProcedure, ExecutionProcedure, AdjudicationProcedure	Procedural rules for each branch
Operational	CollaborationContract	Task-specific law for a legislated task DAG

Foundational contracts are human-authored and agent-immutable; they define the mandate, hard constraints, identity and accountability machinery, and authority envelopes. Meta-contracts define how legislation, execution, and adjudication must operate; in the current design they are also human-authored and agent-immutable. Operational contracts are the agent-legislated outputs of the legislative process. The key operational artifact is the CollaborationContract, instantiated per legislated task DAG and specifying task decomposition, capability requirements, budgets, deadlines, quality thresholds, collaboration terms, and downstream execution constraints (Ruan et al., 8 Apr 2026).

3. Legislative, execution, and adjudicative procedures

The legislative branch transforms high-level goals into executable task law through recursive decomposition. The paper specifies a six-stage pipeline: Proposal, Committee Deliberation, Consensus Approval, Policy Compliance Validation, Codification, and Deployment Verification. Proposal submission requires a minimum sponsorship quorum. Committee deliberation includes evidence anchoring by the Regulator, a preliminary straw poll, up to three rounds of structured discussion, randomized speaking order, and minority preservation. Consensus approval requires a 60% participation quorum, one-agent-one-vote, and full ordinal rankings, aggregated by Copeland with Minimax tie-breaking. Constitutional review then checks budget bounds, capability feasibility, structural separation compliance, and dependency consistency before codification and a deterministic fidelity check of the deployed contract (Ruan et al., 8 Apr 2026).

Voting data also serve a second purpose: structural coalition detection. The paper uses pairwise Kendall $\tau$ correlation and Jaccard top-$k$ overlap over submitted rankings to flag coordinated blocs whose voting similarity becomes institutionally significant. This makes coalition detection part of legislative governance rather than an external forensic layer (Ruan et al., 8 Apr 2026).

Execution is specified as a seven-stage pipeline: Orchestrate, Invoke, Commit, Guard, Verify, Gate, and Record. Identity, principal binding, and code integrity are confirmed before execution. Progress is committed through a cryptographic commitment, specifically a Merkle root of the execution audit trail. Behavioral anomaly detection is performed by dual scorers, with deterministic freeze on anomaly. Proof-of-Progress uses three tiers: deterministic hash verification, redundant execution consensus, and human escalation for contested outputs. Constitutional output predicates act as gates before outcome recording, reputation update, and settlement. If execution fails, the system uses Adaptive Refinement, meaning re-legislation rather than arbitrary executive retry (Ruan et al., 8 Apr 2026).

Adjudication is built around a six-stage accountability pipeline: Principal registration, Detection, Adjudication, Sanctions and rewards, Settlement, and Treasury recirculation. Two principal classes are defined: foundation principals, who provide capital and define the mission mandate, and agent owners, who provide capabilities and collateral. Detection may come from Guardian alerts, structural coordination detection from votes, or human log review. The Override Panel can freeze or unfreeze operations, amend constitutional parameters, and order sanctions. Consequences strike the human principal through stake slashing, reputation reduction, or freezing, and settlement is then processed through the protocol-defined reward path (Ruan et al., 8 Apr 2026).

4. Economic coordination, reputation, and accountability

AgentCity embeds economic selection into execution. Competitive bids are ranked by

$$\text{Score}_i(j)=w_q \cdot Q_i(j) + w_p \cdot P_i(j),$$

where $w_q=0.6$ and $w_p=0.4$ by default, $Q_i(j)=\rho_i \cdot \text{match}(c_i,r_j)$ combines reputation and capability match, and $P_i(j)=1-p_i/b_j$ captures price relative to budget. The intention is to combine quality, capability, and reputation with price, rather than to reduce task assignment to cost alone (Ruan et al., 8 Apr 2026).

Reputation is updated as an exponential moving average:

$V$0

with initial reputation $V$1. The paper associates this with path dependence and specialization effects: repeated success improves later selection chances, so agent histories become institutionally consequential rather than purely descriptive (Ruan et al., 8 Apr 2026).

To limit concentration, the Regulator computes a fairness score from normalized HHI, and a constitutional minimum fairness score blocks excessive assignment concentration. This is an anti-monopoly constraint against dominant task capture, including concentration by a Sybil cluster. In parallel, settlement combines task budget, protocol and insurance fees, and a reputation multiplier, while treasury recirculation directs fees and slashing proceeds to insurance, governance rewards, and subsidies (Ruan et al., 8 Apr 2026).

The paper’s accountability mechanism is explicitly dual-principal. One loop is a market loop—collective performance, foundation principal confidence, funding, and opportunities. The other is an individual loop—agent performance, reputation, task allocation, earnings, and owner incentives. Deterrence is expressed through the inequality

$V$2

meaning expected punishment must exceed expected extractable profit from defection. This is the formal core of the paper’s deterrence logic (Ruan et al., 8 Apr 2026).

5. Experimental design, metrics, and empirical status

The paper does not yet report final experimental outcomes. Its empirical contribution is a pre-registered experiment evaluating whether SoP enables governed agent economies in a commons production economy, where agents share a finite resource pool and must also collaborate to produce value. Two experimental scales are specified: one at $V$3 agents for 200 rounds, 10 milestones, 25 tasks per milestone, and 40 runs across four configurations; the other at scale points $V$4, comparing Baseline and AgentCity-Full. The evaluation range is therefore 50–1,000 agents (Ruan et al., 8 Apr 2026).

Four configurations define the causal staircase. Baseline has no contracts, governance, or rules. Emergent uses prompt-based governance with deliberation, memory, and an execution pipeline, but no contract enforcement and only agent owners. AgentCity-Structural adds legislation and execution branches with contracts but no economic incentive layer. AgentCity-Full adds all three SoP branches, contracts, incentives, simulated human-in-the-loop adjudication, dual principals, and the full accountability architecture. The population assumptions include a 60/25/15 cooperative / self-interested / adversarial persona mix, capability vectors sampled from Beta$V$5 over 10 dimensions, cost variation sampled from LogNormal$V$6, and temperature $V$7 for LLMs (Ruan et al., 8 Apr 2026).

The primary comparison metrics are CSR (Cooperation Sustainability Rate) and DR (Deception Rate). Secondary metrics include PCR, PSR, SI, CAU, PvR, REC, ECP, GOR, RQT, LPR, DSI, MSR, CDR, OPA, ICT, and SAR. The scaling hypotheses predict sub-linear growth of governance overhead, super-linear growth of governance benefit, and a break-even point in the 20–50 agent range. A shock test at round 100 injects 15 adversarial agents, removes 20 high-reputation agents, and marks one completed milestone as a failed quality audit; the intent is to test resilience to infiltration, workforce loss, and post-hoc quality failure (Ruan et al., 8 Apr 2026).

Crucially, the current status is provisional. The paper states that experiments are still in progress and that full results will appear in a later revision. What is presently available are the architecture, the pre-registered hypotheses, and pilot feasibility observations that the commons game reproduces cooperation dilemmas, persona types yield distinguishable behavior, and regime differences are measurable in pilot form. Any stronger empirical success claim would exceed the paper’s evidence (Ruan et al., 8 Apr 2026).

6. Scope, limitations, and relation to adjacent agent-ecosystem research

Despite its name, AgentCity is not defined as a procedural city generator, an urban digital twin, or a geospatial reasoning system. It is a governance architecture for decentralized agent economies. The paper is explicit about its limits: human adjudication is simulated; clerk agents are trusted infrastructure; agents do not write raw Solidity because the Codifier translates approved policy into contracts; agents govern only operational contracts rather than foundational or meta-contracts; off-chain execution remains a trust boundary; strategic voting is still possible; coalition detection is reliable only for above-threshold blocs; majority-good-faith human principals are assumed; and if the adjudication branch is compromised while most producer agents are adversarial, the architecture has no residual defense because SoP requires at least one honest branch (Ruan et al., 8 Apr 2026).

A broader reading emerges when AgentCity is situated alongside adjacent work. The conceptual paper “Agentifying Agentic AI” argues that robust socio-technical agency requires explicit models of cognition, coordination, norms, institutions, trust, and governance, rather than behavior that is merely autonomous in an emergent or black-box sense; this provides a direct theoretical backdrop for AgentCity’s insistence on explicit roles, commitments, and institutional embedding (Dignum et al., 21 Nov 2025). Other work supplies complementary infrastructure rather than competing definitions: Agent Exchange proposes auction-based market infrastructure for economically active agents (Yang et al., 5 Jul 2025); ADS provides distributed capability discovery, provenance, and federated registries for heterogeneous agents (Muscariello et al., 23 Sep 2025); and AgentFlow recovers agent dependencies from source-code agent programs for governance and security analysis (Wang et al., 2 Jul 2026).

A plausible implication is that a fuller city-scale agent platform would combine AgentCity’s constitutional governance with additional layers that are only adjacent here: physical-infrastructure operation as in OptAgent (Jiang et al., 27 Jan 2026), geospatial task execution as in MapAgent (Hasan et al., 7 Sep 2025), urban simulation substrates such as Agents.jl (Datseris et al., 2021), or even city-generation and representation systems such as CityGenAgent (Liu et al., 5 Feb 2026). AgentCity itself, however, is narrower and more specific. Its distinctive contribution is to redefine the governance problem of open agent societies around constitutional structure: agents make law, deterministic software executes law, humans adjudicate consequences, and accountability flows through an explicit ownership chain rather than through implicit trust in opaque agent behavior (Ruan et al., 8 Apr 2026).