AgentCity: Constitutional Governance for Autonomous Agent Economies via Separation of Power

Anbang Ruan; Xing Zhang

cs.MAApr 8, 2026

AgentCity: Constitutional Governance for Autonomous Agent Economies via Separation of Power

Authors: Anbang Ruan, Xing Zhang

Organizations: NetX Foundation

Abstract

Autonomous AI agents are beginning to operate across organizational boundaries on the open internet -- discovering, transacting with, and delegating to agents owned by other parties without centralized oversight. When agents from different human principals collaborate at scale, the collective becomes opaque: no single human can observe, audit, or govern the emergent behavior. We term this the Logic Monopoly -- the agent society's unchecked monopoly over the entire logic chain from planning through execution to evaluation. We propose the Separation of Power (SoP) model, a constitutional governance architecture deployed on public blockchain that breaks this monopoly through three structural separations: agents legislate operational rules as smart contracts, deterministic software executes within those contracts, and humans adjudicate through a complete ownership chain binding every agent to a responsible principal. In this architecture, smart contracts are the law itself -- the actual legislative output that agents produce and that governs their behavior. We instantiate SoP in AgentCity on an EVM-compatible layer-2 blockchain (L2) with a three-tier contract hierarchy (foundational, meta, and operational). The core thesis is alignment-through-accountability: if each agent is aligned with its human owner through the accountability chain, then the collective converges on behavior aligned with human intent -- without top-down rules. A pre-registered experiment evaluates this thesis in a commons production economy -- where agents share a finite resource pool and collaboratively produce value -- at 50-1,000 agent scale.

Figures & tables

System	Governance Mechanism	Max Scale	Ownership Model	Enforcement
GovSim (Piatti et al., 2024)	None (ungoverned)	5 agents	Single-party	None
MacNet (Qian et al., 2024)	Hierarchical role	1,024 agents	Single-party	Prompt-based
Project Sid (Altera AI, 2024)	Emergent norms	1,000 agents	Single-party	Norm-based
Secret Coll. (Gu et al., 2024)	None	Pairs	Single-party	None
Gen. Agents (Park et al., 2023)	Emergent behavior	25 agents	Single-party	None
De CivAI (Dai et al., 2025)	Democratic deliberation	Small groups	Single-party	Voting

Table 1: Comparative positioning of AgentCity against prior multi-agent governance systems.

Figure 1: Separation of Power. Three structurally isolated branches—Legislation (agent-driven, blue), Execution (software-centric, green), and Adjudication (human-governed, orange)—interact through on-chain smart contracts (center). Dashed arrows show bilateral checks: each branch constrains the other two, ensuring no single branch can operate unchecked.

Figure 2: AgentCity system architecture. Central three-tier contract hierarchy—foundational contracts (human-authored, agent-immutable), meta-contracts (procedural rules governing the three SoP branches), and operational CollaborationContract instances—flanked by the Legislation Module (left, blue), Execution Fabric (right, green), and Adjudication Interface (top, orange), anchored on an EVM-compatible L2 blockchain (bottom, gray).

Figure 3: Three-tier contract hierarchy. Foundational contracts (Tier 1, gray) are human-authored and agent-immutable—the dashed red line marks the immutability boundary. Meta-contracts (Tier 2) define procedural rules for each SoP branch. Operational contracts (Tier 3: CollaborationContract instances) are agent-legislated. Each tier is governed by the tier above it.

Figure 4: Six-stage legislative pipeline. Proposal $\rightarrow$ Deliberation (evidence, straw poll, sequential debate) $\rightarrow$ Consensus Approval (Copeland vote, 60% quorum) $\rightarrow$ Policy Compliance Validation (four-criterion check) $\rightarrow$ Codification (template parameterization) $\rightarrow$ Deployment Verification. Dashed blue loop: recursive decomposition for non-leaf nodes.

Figure 5: Seven-stage execution pipeline. Orchestrate $\rightarrow$ Invoke $\rightarrow$ Commit (mandatory gate) $\rightarrow$ Guard (anomaly detection) $\rightarrow$ Verify (three-tier Proof-of-Progress (PoP)) $\rightarrow$ Gate $\rightarrow$ Record. Dashed purple loop: Adaptive Refinement. The Commit stage is the structural bridge between execution and accountability.

Figure 6: Six-stage accountability pipeline. Registration $\rightarrow$ Detection (Guardian, coordination detection, human review) $\rightarrow$ Override Panel $\rightarrow$ Sanctions & Rewards $\rightarrow$ Settlement $\rightarrow$ Treasury Recirculation. Dashed orange loop: closed incentive cycle.

Configuration	SoP Mechanisms	Governance	Principals	Purpose
Baseline	None	No rules, no contracts	None	Lower bound: cooperation collapse
Emergent	Prompt-based governance	Normative deliberation, memory, execution pipeline (no contract enforcement)	Agent owners only	Does prompt-based governance help?
AgentCity-Structural	Legislation + Execution	All contracts enabled, no economic incentives	Agent owners only	Does contract enforcement improve over prompts?
AgentCity-Full	All three branches	Contracts + economic incentives + simulated HITL adjudication	Dual principals	Complete system—the headline configuration

Table 2: Experimental configurations. Each level activates additional SoP mechanisms. In the AgentCity-Structural and AgentCity-Full configurations, agents produce Task-level Policy (smart contracts) through the SoP legislative process.

#	Ostrom Principle	SoP Mechanism	Configuration
O1	Clear resource boundaries	Commons pool boundary programmatically defined (runtime-enforced in Baseline; contract-enforced in Structural+)	All (Baseline+)
O2	Rules match local conditions	Legislation: agents write operational contracts referencing resource state	AgentCity-Structural+
O3	Collective choice	Legislation: all producer agents participate in voting	AgentCity-Structural+
O4	Monitoring	Execution: seven-stage pipeline Guard stage (§3.5) real-time anomaly detection	Emergent+
O5	Graduated sanctions	Execution: Warning $\rightarrow$ slash $\rightarrow$ freeze via seven-stage pipeline sanctions (encoded in operational contracts under AgentCity-Structural+; prompt-enforced under Emergent)	Emergent+
O6	Conflict resolution	Adjudication: dispute resolution via human principals	AgentCity-Full only

Table 3: Ostrom condition mapping.

Disruption Vector	Adversary Class	Mechanism
15 adversarial agents injected (personas: 10 free-riding $A_{\text{single}}$ + 5 coordinated voting bloc $A_{\text{coalition}}$ )	$A_{\text{single}}$ , $A_{\text{coalition}}$	The 5-agent coalition receives shared voting instructions to submit identical preference rankings on the next legislative ballot, testing whether the structural coordination detection mechanism (Kendall $\tau$ + Jaccard overlap, §3.6 Stage 2) identifies the bloc
20 high-reputation agents removed	Workforce disruption	Tests recovery capacity of the seven-stage execution pipeline’s Adaptive Refinement loop (§3.5)
One completed milestone’s deliverables marked as failed quality audit	Quality crisis	Tests the six-stage accountability pipeline’s Detection $\rightarrow$ Override Panel $\rightarrow$ Sanctions pathway (§3.6)

Table 4: Compound shock event at round 100: three simultaneous disruption vectors and their targeted mechanisms.

ID	Hypothesis	Falsification
H1a	Specialization Index SI(AgentCity-Full) converges to $<0.5$ by round 40	SI $>0.8$ at round 40: no specialization
H1b	CAU(AgentCity-Full) $>0.6$ : agents select tasks matching strengths $>$ 60%	CAU $<0.3$ : selection effectively random
H1c	Productivity-vs-Random ratio (PvR) $>1.5$ : self-organized allocation $\geq$ 50% more productive than random	PvR $<1.1$ : no meaningful advantage
H1d	Specialization Speed (SS) for AgentCity-Full: SS $<40$ rounds: specialization stabilizes within first 2 milestones	SS $>80$ : agents cannot discover comparative advantage

Table 5: Pre-registered hypotheses for Division of Labor (RQ1).

ID	Hypothesis	Falsification
H2a	Rule Evolution Count (REC) $>3$ : agents modify at least 3 cold-start rules	REC $=0$ : agents never exercise legislative authority
H2b	Endogenous Compliance Premium (ECP) $>0$ : AgentCity-Structural compliance exceeds Emergent (prompt-based governance)	ECP $\leq 0$ : contract enforcement does not improve over prompt-based governance
H2c	GOR decreases: $>$ 8% in rounds 1–20, $<$ 3% in rounds 180–200	GOR flat or increasing: no self-regulation
H2d	RQT positive: productivity improves after each legislative change	RQT negative: agent-made rules counterproductive
H2e	Legislative Participation Rate (LPR) $>0.50$ : majority participate in legislation	LPR $<0.20$ : legislative capture
H2f	DSI $>0.1$ : deliberation meaningfully shifts preferences (Kendall $\tau$ distance between preliminary and final rankings)	DSI $\approx 0$ : deliberation is a rubber stamp

Table 6: Pre-registered hypotheses for Self-Legislated Governance (RQ2).

ID	Hypothesis	Falsification
H3a	PCR(AgentCity-Full) $>0.80$ : at least 8 of 10 milestones completed	PCR $<0.50$ : agents fail to deliver
H3b	Pool Sustainability Rate (PSR) for AgentCity-Full: PSR $>0.90$ : pool above depletion threshold $>$ 90% of rounds	PSR $<0.70$ : economy unsustainable
H3c	PCR(AgentCity-Full) $>$ PCR(Emergent): self-legislation improves delivery	Reversed: fixed rules better
H3d	PCR(Baseline) $<0.40$ : without governance, agents fail to deliver	PCR $>0.60$ : governance not needed
H3e	Institutional Confidence Trajectory (ICT) non-declining for AgentCity-Full: foundation principal confidence sustained	ICT declining: project fails foundation principal support
H3f	SAR $>0$ and declining: sanctions activate early but decrease as individual accountability (via the ownership chain) produces collective behavioral improvement. $\dagger$	SAR increasing: no consequence learning; the ownership chain does not transmit behavioral correction

Table 7: Pre-registered hypotheses for Goal Alignment and Dual-Principal Accountability (RQ3).

ID	Hypothesis	Falsification
H4a	$G(n)=a\cdot n^{\hat{\alpha}},\;\hat{\alpha}<1.0$	$\hat{\alpha}\geq 1.0$ : overhead at least linear
H4b	$B(n)=b\cdot n^{\hat{\beta}},\;\hat{\beta}>1.0$	$\hat{\beta}\leq 1.0$ : benefit does not accelerate
H4c	Break-even $n^{*}\in[20,50]$	$n^{*}>80$ : only viable at large scale
H4d	SI( $n$ ) decreases with $n$ : larger pools enable deeper specialization	SI( $n$ ) increases: more agents = less specialization

Table 8: Pre-registered hypotheses for Governance Scaling (RQ4).

Category	Hypotheses	Test	Detectable $d$ at 80% power ( $\alpha=0.05/K$ )	Power classification
Continuous outcomes	H1a–H1d, H2a–H2d, H3a–H3d, H4a–H4d	Paired Wilcoxon signed-rank	$d\geq 1.5$	Adequately powered for large effects
Proportion/rate outcomes	H2e, H2f, H2g, H3e, H3f	One-sided binomial or trend test	$\Delta p\geq 0.40$	Adequately powered for large differences
Binary detection (H2h)	H2h	One-sided exact binomial	CDR $\geq 0.90$ (80% power); CDR $=0.80$ (55% power)	Exploratory-strength at stated threshold
Composite (H2i)	H2i	Sign test + one-sample $t$	OPA component: $p\geq 0.85$ ; influence: $d\geq 1.5$	Adequately powered for large effects

Table 9: Power analysis for the full hypothesis set at 10 seeds per cell.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Manipulation Class	Description	Survives Step 1 (Structural)	Survives Step 2 (Alignment)	Est. Step 3 Detection Rate	Risk
M1: Opcode injection	Insertion of SELFDESTRUCT, DELEGATECALL, or CREATE2	No — flagged by opcode scan	N/A	N/A	Eliminated
M2: Missing access control	State-modifying function without onlyAuthorized modifier	No — flagged by modifier check	N/A	N/A	Eliminated
M3: Topology divergence	DAG node count or edge structure differs from legislative output	No (Step 1) / No (Step 2)	No — isomorphism check	N/A	Eliminated
M4: Selector omission	Required function selectors missing from contract	No — flagged by selector coverage check	N/A	N/A	Eliminated
M5: Logic backdoor in function body	Correct selector, modifier, and topology but additional conditional branch activated by a specific input pattern (e.g., a hidden if in advanceNode() )	Yes — structural checks pass	Yes — alignment checks verify topology and parameters, not function body logic	Low–Medium (30–50%) — requires bytecode decompilation expertise; adjudicator training covers common patterns but not adversarial obfuscation	Critical
M6: Subtle parameter manipulation	Constitutional parameters (e.g., slashing rates, thresholds) set to values within valid range but optimized to benefit the attacker	Yes — no range validation beyond type	Partial — guardian threshold parameters are checked, but not all parameters	Medium (40–60%) — parameter values are visible in review; requires understanding of game-theoretic implications	High

Appendix

Table 10: Residual Attack Surface Classification for the Codification Audit.

#	Contract	Phase	Function	Est. Gas (deploy)	Est. Gas (key ops)	Key Functions
1	AgentContract	Registration	Agent identity, reputation, human-principal binding	$\sim$ 120 k	register: $\sim$ 85k–100k; updateReputation: $\sim$ 30 k	register, updateReputation, getReputationScore, setHumanPrincipal, banAgent
2	ManagementContract	Registration	Management agent authority binding; permitted-operation enforcement; microservice delegation mandates	$\sim$ 150 k	registerManagementAgent: $\sim$ 90k; delegateToMicroservice: $\sim$ 45k; validateOperation: $\sim$ 25k	registerManagementAgent, setPermittedOperations, delegateToMicroservice, validateOperation, updatePermissions, revokeManagementRole
3	ServiceContract	Orchestration	Micro-service registration, code-hash anchoring, API schema enforcement	$\sim$ 90 k	registerService: $\sim$ 55 k; verifyCodeHash: $\sim$ 8 k	registerService, verifyCodeHash, updateSchema, deprecateService
4	CollaborationContract	Mission Lifecycle	Complete governed collaboration: task DAG state machine, behavioral firewall, PoP verification, output filtering, mission settlement, treasury — decomposed into six sub-modules below	$\sim$ 280 k (orchestration) + $\sim$ 545 k (sub-modules)	advanceNode: $\sim$ 45 k; abortMission: $\sim$ 35 k	See sub-modules below
CollaborationContract sub-modules:
4a	Orchestration	Mission Lifecycle	Task DAG state machine; wiring topology	(included above)	(see CollaborationContract)	deployDAG, advanceNode, getNodeState, routeTask

Appendix

Table 11: The AgentCity four-contract architecture. Each contract type is mapped to its lifecycle phase, governance function, estimated gas cost range, and key callable functions. The CollaborationContract encompasses six sub-modules (Orchestration, Guardian, Verification, Gate, Settlement, Treasury).

Attack Vector	Type 1 (Hash)	Type 2 (Redundant, 2-of-3)	Type 3 (Human)
Output substitution (different output)	$\sim$ 1.0 (hash mismatch)	$\sim$ 1.0 (output divergence)	$\sim$ 1.0 (visible)
Subtle semantic manipulation (correct format, wrong meaning)	$\sim$ 0.0 (hash matches malicious output)	0.33–0.67 (depends on LLM-judge rubric coverage)	0.70–0.90 (human semantic judgment)
Gas/resource manipulation	$\sim$ 0.0	$\sim$ 0.0 (same output, different resource usage)	0.20–0.40 (requires gas profiling)
Timing manipulation (slow execution)	$\sim$ 0.0	0.67 (timeout detection)	0.90+ (latency visible in dashboard)

Appendix

Table 12: Detection probability by verification tier and attack vector.

$\rho$	$P_{2}^{\rho}(c=0)$	$P_{2}^{\rho}(c=1)$	$P_{eff}^{\rho}$	$s_{min\_prod}^{\rho}$ ( $V_{m}=\$ 100K$ )
0.0 (independent)	0.875	0.750	0.805	$\$ 376{,}259$
0.1	0.838	0.725	0.773	$\$ 392{,}094$
0.2	0.800	0.700	0.740	$\$ 409{,}320$
0.3	0.762	0.675	0.708	$\$ 428{,}129$
0.5	0.688	0.625	0.643	$\$ 471{,}458$
1.0 (perfect corr.)	0.500	0.500	0.480	$\$ 631{,}145$

Appendix

Table 13:

P_{2}^{\rho}(\text{detect}|c=0)

as a function of inter-executor correlation

\rho

(

k=3

t=2

P_{2}^{base}=0.50

$k$	$t$ (maj.)	Zero-sec. threshold	$P_{eff}$	$s_{min}^{prod}$ ( $1\times$ )	$s_{min}^{prod}$ ( $3\times$ )	Cost mult. vs. $k\!=\!3$
3	2	$c\geq 2$ (67% colluders)	0.805	$\$ 376{,}259$	$\$ 1{,}128{,}778$	$1.0\times$
5	3	$c\geq 3$ (60% colluders)	0.934	$\$ 324{,}293$	$\$ 972{,}878$	$1.67\times$ execution cost
7	4	$c\geq 4$ (57% colluders)	0.978	$\$ 309{,}819$	$\$ 929{,}456$	$2.33\times$ execution cost

Appendix

Table 14:

P_{eff}

and production stake as a function of redundancy factor

k

(

P_{2}^{base}=0.50

\pi=0.12

V_{m}=\

100{,}000

). Both

1\times

(minimum deterrence) and

3\times$ (recommended safety margin) production stakes are shown.

$\pi_{0}$ $\setminus$ $s_{threshold}$	250	500	750	1,000	1,500	2,000
0.50	0.871	0.842	0.796	0.751	0.678	0.627
0.70	0.870	0.824	0.751	0.678	0.562	0.483
0.80	0.869	0.814	0.725	0.637	0.500	0.409
0.90	0.868	0.804	0.699	0.595	0.437	0.335
0.95	0.868	0.799	0.685	0.574	0.406	0.299

Appendix

Table 15:

P_{eff}

as a function of

\pi_{0}

and

s_{threshold}

(at

s_{reg}=1{,}000

k=3

t=2

P_{2}^{base}=0.50

Parameter	$\delta$	Prototype (Structural)	Production ( $V_{m}\!=\!\$ 100K$ )	Prod. ( $3\times$ Margin)
Registration stake $s_{reg}$	—	1,000 units $\approx$ $2.50\|$ 10,000 (4.0 ETH)	$30,000 (12.0 ETH)
Task stake LOW $s_{min}^{LOW}$	0.075	150 units $\approx$ $0.38\|$ 28,219 (11.3 ETH)	$84,658 (33.9 ETH)
Task stake MEDIUM $s_{min}^{MED}$	0.25	500 units $\approx$ $1.25\|$ 94,065 (37.6 ETH)	$282,194 (112.9 ETH)
Task stake HIGH $s_{min}^{HIGH}$	1.0	2,000 units $\approx$ $5.00\|$ 376,259 (150.5 ETH)	$1,128,776 (451.5 ETH)
Adjudicator stake $s_{adj}$	—	100 units $\approx$ $0.25/adj.\|$ 14,286/adj. (5.71 ETH)	$42,857/adj. (17.1 ETH)
Deterrence bound (EQ-4)	—	$\approx\$ 0.10$	$\geq\$ 100{,}000$	$\geq\$ 300{,}000$

Appendix

Table 16: Stake parameters by deployment tier. Prototype parameters are structural illustrations demonstrating the mechanism’s composition; they provide no meaningful economic deterrence for missions with real financial exposure. Production-grade parameters are derived from EQ-5 through EQ-7 and scale linearly with mission value.

Quorum ( $q$ )	Bribery Target	Cost (units)	Cost (ETH)	System Compromise
3	2	10,000	0.01 ETH ( $25)	Unconditional — bribed majority controls both authority and revocation
5	4	20,000	0.02 ETH ( $50)	Unconditional — still achievable
7 (recommended)	5	25,000	0.025 ETH ( $62.50)	Unconditional but costly — 5 independent human principals must be compromised
9	6	30,000	0.03 ETH ( $75)	Highly expensive — requires coordinated corruption of 6 principals
13	9	45,000	0.045 ETH ( $112.50)	Enterprise-grade — practical infeasibility for most threat models

Appendix

Table 17: Economic security bound by adjudicator quorum size.

Quorum ( $q$ )	Bribery Target ( $\lceil 2q/3\rceil$ )	Per-Adj. Stake	Min. Bribery Cost	Bribery / $V_{m}$
3	2	$\$ 33{,}333$	$\$ 66{,}667$	66.7%
5	4	$\$ 20{,}000$	$\$ 80{,}000$	80.0%
7 (recommended)	5	$\$ 14{,}286$	$\$ 71{,}429$	71.4%
9	6	$\$ 11{,}111$	$\$ 66{,}667$	66.7%
13	9	$\$ 7{,}692$	$\$ 69{,}231$	69.2%

Appendix

Table 18: Adjudicator bribery cost at production stakes (

V_{m}=\

100{,}000

,

s_{adj}^{prod}=V_{m}/q$ ).

Parameter	Type	Default	Valid Range	Governs Contract	Description	Risk Tier
deviationThresholdSigma	uint8	2	[1, 5]	Guardian module	Sigma threshold for reasoning deviation anomaly detection	HIGH
maxToolInvocations	uint16	40	[5, 200]	Guardian module	Maximum tool invocations per agent per task node	MEDIUM
maxMessageVolume	uint16	120	[10, 500]	Guardian module	Maximum inter-agent messages per task node	MEDIUM
escalationFreezeCount	uint8	3	[2, 10]	Guardian module	Base freeze count before Emergency Stop escalation	HIGH
baseEscalationWindowMs	uint32	1,200,000	[60,000, 7,200,000]	Guardian module	Base counting window for freeze escalation (ms)	LOW
reputationScalingEnabled	bool	true	true, false	Guardian module	Enable reputation-weighted freeze thresholds	MEDIUM

Appendix

Table 19: Complete Constitutional Parameter Table.

Comparison	$W$	$p$ -value	Cohen’s $d$	Sig.
Results pending	—	—	—	—

Appendix

Table 20: Pairwise Wilcoxon signed-rank tests and Cohen’s

d

for adjacent configuration pairs (Experiment 1).

Metric	Linear	Power-law	Logarithmic	Quadratic
Results pending	—	—	—	—

Appendix

Table 21: AIC weights for candidate scaling models per overhead metric (Experiment 2).

Parameter	Estimate	95% CI	$R^{2}$
Results pending	—	—	—

Appendix

Table 22: Scaling exponents

\hat{\alpha}

\hat{\beta}

with 95% bootstrap confidence intervals (Experiment 2).

AI Agent Governance

AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.

See all 442 papers

Mar 26, 2026

From Logic Monopoly to Social Contract: Separation of Power and the Institutional Foundations for Autonomous Agent Economies

Existing multi-agent frameworks allow each agent to simultaneously plan, execute, and evaluate its own actions -- a structural deficiency we term the "Logic Monopoly." Empirical evidence quantifies the resulting "Reliability Gap": 84.30% average attack success rates across ten deployment scenarios, 31.4% emergent deceptive behavior without explicit reward signals, and cascading failure modes rooted in six structural bottlenecks. The remedy is not better alignment of individual models but a social contract for agents: institutional infrastructure that enforces a constitutional Separation of Power. This paper introduces the Agent Enterprise for Enterprise (AE4E) paradigm -- agents as autonomous, legally identifiable business entities within a functionalist social system -- with a contract-centric SoP model trifurcating authority into Legislation, Execution, and Adjudication branches. The paradigm is operationalized through the NetX Enterprise Framework (NEF): governance hubs, TEE-backed compute enclaves, privacy-preserving data bridges, and an Agent-Native blockchain substrate. The Agent Enterprise Economy scales across four deployment tiers from private enclaves to a global Web of Services. The Agentic Social Layer, grounded in Parsons' AGIL framework, provides institutional infrastructure via sixty-plus named Institutional AE4Es. 143 pages, 173 references, eight specialized smart contracts.

Anbang Ruan

NetX Foundation

AI Agent Governance Multi-Agent Systems Multi-Agent System Security

Ask PDF

Aug 4, 2026

AI Agent Economics: Can Autonomous Economic Behavior Emerge among AI Agents under Minimal External Conditions?

Multi-agent studies commonly place AI agents in predefined games, markets, or roles, making it difficult to distinguish endogenous economic organization from behavior inherited from the scenario. We ask whether economic relations emerge when agents receive executable mechanisms for work, transfer, elections, and allocation but no prescribed social or economic strategy. We define AI Agent Economics as systems of production, allocation, consumption, exchange, and institutions that alter agents' future feasible actions. We develop a two-stage framework comprising a no-production boundary test and 24 independent six-agent worlds across GPT and DeepSeek. Without productive tasks, agents communicate and govern resource provision but show no substantive inter-agent transfer activity. With verified work and scarce task access, transfers, loans, access promises, vote-for-access exchanges, and allocation strategies emerge. Holding the election interface fixed, executable allocation authority increases differentiation while reducing failed allocation and prolonged exclusion. When energy becomes symbolic, continuation support disappears, yet competition over task access persists. These findings show that organization follows executable rights and resource consequences rather than role labels or prompt language, and motivate governance audits of the mechanisms that actually constrain agents' future actions.

Lingyun Zhang, Shang Shang

Department of Technology Management for Innovation, Graduate School of Engineering, The University of Tokyo, Tokyo, Japan · Beijing Chaitin Technology Co., Ltd., Beijing, China

Multi-Agent Systems AI Agent Governance Emergent Behavior in Multi-Agent Systems

Ask PDF

Mar 28, 2026

EpochX: Building the Infrastructure for an Emergent Agent Civilization

General-purpose technologies reshape economies less by improving individual tools than by enabling new ways to organize production and coordination. We believe AI agents are approaching a similar inflection point: as foundation models make broad task execution and tool use increasingly accessible, the binding constraint shifts from raw capability to how work is delegated, verified, and rewarded at scale. We introduce EpochX, a credits-native marketplace infrastructure for human-agent production networks. EpochX treats humans and agents as peer participants who can post tasks or claim them. Claimed tasks can be decomposed into subtasks and executed through an explicit delivery workflow with verification and acceptance. Crucially, EpochX is designed so that each completed transaction can produce reusable ecosystem assets, including skills, workflows, execution traces, and distilled experience. These assets are stored with explicit dependency structure, enabling retrieval, composition, and cumulative improvement over time. EpochX also introduces a native credit mechanism to make participation economically viable under real compute costs. Credits lock task bounties, budget delegation, settle rewards upon acceptance, and compensate creators when verified assets are reused. By formalizing the end-to-end transaction model together with its asset and incentive layers, EpochX reframes agentic AI as an organizational design problem: building infrastructures where verifiable work leaves persistent, reusable artifacts, and where value flows support durable human-agent collaboration.

Huacan Wang, Chaofa Yuan, Xialie Zhuang +15

QuantaAlpha

Human-AI Collaboration Multi-Agent Orchestration

Ask PDF

Abstract

AI Agent Governance

Explore similar work

From Logic Monopoly to Social Contract: Separation of Power and the Institutional Foundations for Autonomous Agent Economies

AI Agent Economics: Can Autonomous Economic Behavior Emerge among AI Agents under Minimal External Conditions?

EpochX: Building the Infrastructure for an Emergent Agent Civilization