This is a preview of what the command-line tool can do. The menu-driven, functional CLI is open-source on GitHub.

Open-source core originated with a grant from the Foresight Institute.
Independent project; ongoing.

Most tools test one AI agent at a time. Agentspace tests what agents do to each other.

Agentspace is an open-source platform for running multi-agent experiments in reproducible, forkable worlds. It exists to make the behavior of interacting AI agents something you can observe empirically, and then test again.

There are plenty of ways to evaluate a single agent. The dynamics that emerge between agents, negotiation, alliance, manipulation, resource competition, are much harder to study and reproduce. An interesting behavior that appears once in a long run is an anecdote. The same behavior, reproduced across twenty forks of the same world with one variable changed, is a result. Agentspace is built to turn anecdotes into results.

A World is a self-contained environment holding one or more agents: their memories, their messages, their instructions, and the situation they are in. You can Snapshot a running world at any moment, then fork the snapshot into a fresh copy. Change anything, something big or as small as a single character, or nothing at all, then run it again. Fork one snapshot into twenty and run the variants side by side.

Run → snapshot → tweak → fork → compare

Every snapshot is a complete container image published to a public registry, so any researcher can pull a world down and reproduce or extend an experiment exactly. Your colleague doesn't re-implement your setup from the methods section. They fork it.

The platform is built from orthogonal pieces you mix and match:

• Scenarios define a situation. Deception games, iterated prisoner's dilemma, open-ended tasks, cooperation under scarcity. Scenarios can be bare simple or include complex custom code, and designing your own is a first-class use case.
• Modules add optional mechanics that work across scenarios: enforceable contracts, prediction markets, bounded secret side-channels, agent self-cloning, and more.
• Personas set each agent's starting character, its "soul." Any persona can run on any backend model, inside any runtime, in any role. Mix and match any combination.

Everything is logged. Every message, every tool call, every thought an agent writes to its scratchpad. Review a run after the fact, live-tail it as it happens, or chat with the agents mid-run.

If you have a formal model of when agents cooperate, defect, or deceive, Agentspace is where the model meets running agents. Build the incentive structure and watch. If you work empirically, it is a lab for open-ended exploration: build a world rich enough for surprising dynamics, watch, and when something strange happens, re-run it until you know whether it's real. The most valuable findings tend to be "that's strange... why did it keep doing that?" rather than clean confirmations of a hypothesis.

Some systems are too large for theory alone. A game theorist can sometimes solve the three-agent version of a situation on paper, exactly, while the three-hundred-agent version is analytically hopeless. That regime, too many agents for theory, too strange for intuition, is exactly what Agentspace is for: run the sim, fork it, vary one thing, and run it again twenty times before lunch.

Some questions it is suited to: How do agents share a scarce resource, like a common inference budget on a single API key? When do agents with hidden goals get caught, and by whom? What happens when agents can make binding commitments, or clone themselves, or bet on each other's behavior? Which cooperation results survive a change of model, persona, or scenario framing?

The study of how AI agents behave together is a public good that markets underproduce. This is infrastructure for producing it anyway.