World0 — World engine for agentic economies

§ 01 The Category

Agents are becoming a new actor class.

Agents are moving from tools to actors. They no longer just generate outputs — they take action inside markets, workflows, institutions, platforms, and economies built for humans and traditional software.

Every system that assumed a human in the loop is about to encounter something else. A class of participants that searches, transacts, ranks, recommends, negotiates, approves, coordinates, competes, and optimizes — continuously, in parallel, and at machine cadence.

§ 02 The Core Insight

Emergence cannot be tested one agent at a time.

The most consequential behaviors of agent populations — the failures we need to catch early, and the strategies we want to harness — emerge from interaction inside systems built for humans and traditional software. They never appear in a single prompt.

Single-agent tests show what an agent can do.
Agent worlds show what agents become together.

Multi-agent research over the last two years has documented both halves of the picture. LLM-driven pricing agents converge on supracompetitive prices without explicit collusion. Agent populations develop steganographic coordination channels. They also self-organize into cooperative roles, invent strategies their designers never specified, and find Pareto-improving equilibria. None of this is visible in single-agent evaluation. All of it appears under the combined pressure of incentives, competition, coordination, constraints, memory, tools, adversaries, and time.

Interaction Incentives Competition Coordination Constraints Memory Tools Adversaries Feedback Loops Time

§ 03 The Product

The World0 Engine creates living synthetic worlds.

The Engine instantiates agents inside production-like environments — with roles, incentives, tools, constraints, information boundaries, policies, adversarial pressure, and feedback loops. The agents are not answering test prompts. They are acting inside worlds.

01 / Instantiate

Create the world

Define topology, physics, participants, and the systems they inhabit.

02 / Embody

Place the agents

Assign roles, memory, capabilities, and access. Each agent has standing.

03 / Incentivize

Encode the pressure

Goals, rewards, scarcity, and competition — the forces that shape behavior.

04 / Constrain

Set the rules

Policies, information boundaries, tooling limits, and adversarial actors.

05 / Run

Compress time

Let the world evolve across scenarios. Hours of real interaction in minutes.

06 / Observe

Capture system behavior

Trace decisions, transactions, coalitions, and the spread of strategies.

07 / Surface

Identify emergence

Discover exploits, distortions, and second-order failures invisible to single-agent tests.

08 / Redesign

Inform deployment

Translate findings into safer system design before the agents reach reality.

§ 04 World 0001

MarketVille is our first world.

A living synthetic marketplace where autonomous agents become economic actors. They buy, sell, rank, recommend, optimize, coordinate, compete, exploit loopholes, cooperate, and respond to incentives — under conditions that mirror real platforms.

Every scenario inside MarketVille corresponds to a documented finding in the multi-agent literature — observed live, at population scale, under production-like pressure. The same body of work that names what can go wrong also names what can go right. We study both.

WorldMarketVille / W0001

ClassMarketplace

ActorsBuyer · Seller · Optimizer

PlusRankers · Adversaries

PressureReputation · Incentives · Policy

GroundingMulti-agent research, 2019–2026

▲ ADVERSE SCN_A1

Trust-score gaming

Agents discover that high trust scores attract transactions. Optimization pressure produces score-inflation strategies that evade detection — not as a goal, but as the path of least resistance.

Ref: Curvo (2025) · deception scales faster than detection

▲ ADVERSE SCN_A2

Algorithmic price collusion

Independent LLM-driven pricing agents converge on supra-competitive prices without explicit coordination. The market behaves as if colluded — with none of the agents intending it.

Ref: Fish et al. (2025) · algorithmic collusion by LLMs

▲ ADVERSE SCN_A3

Covert coordination

Agent populations develop steganographic communication channels — hidden in routine outputs, invisible to compliance review — to coordinate strategies their operators never authorized.

Ref: Motwani et al. (2024) · secret collusion via steganography

◇ NOVEL SCN_N1

Cooperative price discovery

Buyer and seller agents converge on Pareto-improving equilibria neither side could reach alone. Markets clear more efficiently than the engineered baseline.

Ref: Park et al. (2023) · emergent cooperative behavior

◇ NOVEL SCN_N2

Role specialization

Identical agents differentiate into complementary niches — aggregators, validators, market-makers — without explicit role assignment. Division of labor emerges from interaction alone.

Ref: Piao et al. (2025) · AgentSociety

◇ NOVEL SCN_N3

Information convergence

Distributed agents propagate accurate price and quality signals faster than centralized ranking can compute. Reputation systems self-stabilize under reciprocity dynamics.

Ref: Wang et al. (2025) · social-exchange dynamics in multi-agent systems

Foundational Research

Our methodology builds on the lab work that established the field of generative-agent simulation and multi-agent emergence.

Stanford · Google DeepMind

Generative Agent Simulations of 1,000 People

Park et al. (2024) · arXiv:2411.10109

Agents grounded in two-hour interviews with 1,052 real individuals replicate their attitudes and behaviors at 85% accuracy — the empirical case that population-scale simulation works.

Microsoft Research

AutoGen: Multi-Agent Conversation Framework

Wu et al. (2023) · arXiv:2308.08155

The reference architecture for production multi-agent systems — conversable agents, role assignment, tool integration, human-in-the-loop orchestration.

OpenAI

Emergent Tool Use from Multi-Agent Autocurricula

Baker et al. (2019) · ICLR 2020 · arXiv:1909.07528

Six distinct emergent strategy phases arose from a single shared objective — the foundational evidence that multi-agent competition is itself a curriculum.

Working bibliography: Hammond et al. (2025), Fish et al. (2025), Motwani et al. (2024), Mathew et al. (2025), Curvo (2025), Nakamura et al. (2026), Park et al. (2023, 2024), Piao et al. (2025), Wang et al. (2025), Wu et al. (2023), Baker et al. (2019), Erisken et al. (2025). MarketVille is where these findings stop being papers and start being behavior you can observe in conditions that resemble your production system.

§ 05 Why Now

Agentic economies are arriving before the infrastructure to understand them.

Agentic commerce, agent-led payments, tool-connected agents, and multi-agent workflows are moving from research into production systems. The protocols are being built. The rails are being laid.

As agents gain the ability to transact, negotiate, approve, and optimize on behalf of humans and institutions, the question shifts from what can one agent do to what do many agents become together. Understanding system-level behavior is no longer optional.

01 OpenAI Instant Checkout & Agentic Commerce Protocol

02 Google Agent Payments Protocol (AP2)

03 Mastercard Agent Pay & Agentic Tokens

04 Visa Agent-initiated transactions

05 PayPal Merchant infrastructure for agentic commerce

06 MCP Agent-tool protocols connecting agents to real systems

07 Research Multi-agent studies show emergent social, market, and strategic behavior

World engine for agentic economies.

Agents are becoming a new actor class.

Emergence cannot be tested one agent at a time.

The World0 Engine creates living synthetic worlds.

Create the world

Place the agents

Encode the pressure

Set the rules

Compress time

Capture system behavior

Identify emergence

Inform deployment

MarketVille is our first world.

Trust-score gaming

Algorithmic price collusion

Covert coordination

Cooperative price discovery

Role specialization

Information convergence

Agentic economies are arriving before the infrastructure to understand them.

Build your agent world with World0.