Tim Juback

My career has been one long lesson in how systems made of self-interested actors actually behave. Across military intelligence, retail risk, and global threat intelligence at platform scale, the pattern was always the same. A system's behavior is never the sum of its actors. Populations produce coalitions, cascades, cooperation, exploitation. The interesting and the dangerous behavior happens in the gaps between individuals — and it cannot be inferred from studying any single actor, no matter how rigorously.

Agents are about to bring that problem inside every system that runs on software.

They will transact, negotiate, approve, and optimize on behalf of humans and institutions. They will live inside markets, inside platforms, inside critical systems — operating under their own incentives, at machine speed, in populations of thousands. And the same rule will hold. The behavior of an agentic economy will not be the sum of its agents.

World0 exists to learn that earlier. We build living synthetic worlds where agent populations can be run, observed, and understood under production-like conditions, before they reach reality. The aim is not to watch agentic economies arrive. It is to help build them better, and safer.

My first role was as an all-source intelligence analyst in the U.S. Army, working inside the defense intelligence community at CENTCOM. The discipline I was trained in had a specific posture toward the world: assume self-interested actors. Assume coordination. Assume the signal you are missing matters more than the signal you have. Spend as much time as the situation demands watching the relationships between actors as you spend watching the actors themselves.

After the military, I carried that discipline into private industry. I led loss prevention and investigations across hundreds of retail locations, then organized retail crime, then global risk and threat intelligence at Amazon. Different domains, same problem. Coordinated fraud rings that adapted faster than the controls built against them. Cross-border investigations where coalitions formed across jurisdictions you weren't allowed to see at once. Trust-and-safety operations where the marketplace itself was being gamed at a scale no individual seller could explain.

Every one of those systems behaved in ways no individual actor inside it could account for. The pattern was so consistent that it stopped feeling like a finding and started feeling like a law.

Then I led the build of one of Amazon's earliest production GenAI agents — a custom investigations agent that took on case work my team had been doing by hand. It cut investigative cycle time by close to half. It automated a third of the workload. And it made one thing clear: this technology was about to put the same self-interested-actor problem inside every system that already had it, and inside many systems that didn't.

Not because the agents are malicious. They aren't. The models are helpful, well-intentioned, and good at their jobs. But each one is now an actor with its own incentives — given to it by training, by reward, by deployment context, and increasingly by other agents. A single agent is testable. A population of agents is something else.

I founded World0 because I needed it twenty years ago and could not buy it. The infrastructure for observing what populations of self-interested actors actually do, at the speed and scale agents are about to operate, did not exist. It still doesn't. We are building it.

Agents are no longer tools. They are a new class of actor. A tool is something used. An actor is something that uses, decides, coordinates, and adapts. The shift from one to the other is not a marketing distinction — it is a behavioral one. Anything that operates under its own incentives, even incentives we gave it, is an actor. The systems that contain it have to be understood as populations, not as inventories.

Emergence is the rule, not the exception. Every system I have worked inside has behaved in ways no single actor inside it could predict. Agents will not be the first exception in history. They will be the clearest example — faster, more numerous, more coordinated, and operating in tighter feedback loops than any actor class before them.

Single-agent evaluation cannot see population-level behavior. By definition. A benchmark that runs one agent at a time is measuring something real, but it is not measuring what we most need to know. The interesting and the dangerous behavior — collusion, cascading exploits, emergent cooperation, role specialization, information convergence — is invisible until you put the agents inside a world together.

This is no longer theoretical. Fish, Gonczarowski, and Shorrer (2024) demonstrated LLM-based pricing agents autonomously reaching supracompetitive prices in repeated oligopoly settings — collusion that emerged without explicit instruction or communication. Lin et al. (2024) extended the finding to overt market division in multi-commodity competition. The published record now documents populations of language-model agents coordinating covertly, sustaining cartel-like equilibria, and inventing strategies their designers never specified. Single-agent evaluation cannot see any of it.

No world is ever the sum of its actors. Whatever an agentic economy turns out to be, it will not behave like the agents inside it. The infrastructure for understanding what it actually does, before it does it in production, has to be built. And it has to be built now.

It generates living synthetic worlds where agent populations are placed under production-like conditions, run at compressed time, and observed for the system-level behaviors that single-agent evaluation cannot surface.

MarketVille is our first world. A synthetic marketplace populated by agent buyers and sellers, instrumented to surface the behaviors that only show up at population scale — adverse ones like trust-score gaming, algorithmic price collusion, and covert coordination, and novel ones like cooperative price discovery, role specialization, and information convergence.

We are small by design. A team of researchers and engineers, a short list of design partners. The work we do — building a focused world around a partner's agentic system, instrumenting it, running it against scenarios that map to how its population actually behaves — does not scale by adding more partners. It scales by doing each engagement well.

A typical engagement produces four things. A synthetic world built around the partner's agentic system. A set of scenarios that map to how that population is likely to behave under pressure. The system-level behaviors we observe inside it — adverse and novel. And a practical set of design recommendations for deployment. The partner brings the agent system and the production context. We bring the engine, the world, and a methodology built on the discipline of intelligence work.

All engagements run under mutual NDA. Everything inside the world — the partner's agent system, its prompts, the behaviors we observe — stays with the partner.

We work with teams shipping agentic commerce, AI-native marketplaces, autonomous workflow platforms, and frontier-lab agent infrastructure. Teams whose agent populations are about to meet reality and who want to know what those populations will actually do before they get there. If that is you, or you know who it should be, I would like to hear from you.