A reproducible testbed for AI agents that interact.
“Empirical work on the safety of multi-principal, multi-agent deployments is bottlenecked by the need for realistic and reproducible testbeds.”
The #1 request in DeepMind's new $10M multi-agent AI safety funding call
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 cooperation and alignment of interacting AI agents something you can observe and test empirically.
Systems to study how agents behave together are a public good that markets currently underproduce. Current incentives lead to building capable single agents for profit. Studying multi-agent cooperation is of academic interest. But that is rapidly changing.
And research is just the beachhead. Big companies want this for the same reasons the safety field does, interacting agents fail in ways you can't see coming, and the only defense is to make that behavior observable and reproducible before it reaches the world. A commercially successful Agentspace means more multi-agent systems get tested for collusion, manipulation, and runaway coordination before they ship. Incentives point the same way whoever the buyer is.
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. You can then change anything about the world, something big or as small as a single character, or nothing at all, then run it again. Run → snapshot → tweak → fork → compare is the whole loop.
The platform is built from orthogonal pieces you mix and match:
Everything is logged. Every message, every tool call, every thought an agent writes to its scratchpad, so you can review a run afterward or live-tail it as it happens. The most valuable findings tend to be “that’s strange... why did it keep doing that?”, and agentspace is built to surface those moments and let you re-run them to test whether the effect is real.
This is a preview, a demo of what the command line tool can do. The menu driven, functional command-line tool is open-source on GitHub.
github.com/sfgeekgit/agentspace →Open-source core originated with a short term grant from the Foresight Institute.
Independent project; ongoing.
This is a preview, a demo of what the command line tool can do. The menu driven, functional command-line tool is open-source on GitHub.
github.com/sfgeekgit/agentspace →Open-source core originated with a short term grant from the Foresight Institute.
Independent project; ongoing.
A reproducible testbed for AI agents that interact.
“Empirical work on the safety of multi-principal, multi-agent deployments is bottlenecked by the need for realistic and reproducible testbeds.”
The #1 request in DeepMind's new $10M multi-agent AI safety funding call
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 cooperation and alignment of interacting AI agents something you can observe and test empirically.
Systems to study how agents behave together are a public good that markets currently underproduce. Current incentives lead to building capable single agents for profit. Studying multi-agent cooperation is of academic interest. But that is rapidly changing.
And research is just the beachhead. Big companies want this for the same reasons the safety field does, interacting agents fail in ways you can't see coming, and the only defense is to make that behavior observable and reproducible before it reaches the world. A commercially successful Agentspace means more multi-agent systems get tested for collusion, manipulation, and runaway coordination before they ship. Incentives point the same way whoever the buyer is.
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. You can then change anything about the world, something big or as small as a single character, or nothing at all, then run it again. Run → snapshot → tweak → fork → compare is the whole loop.
The platform is built from orthogonal pieces you mix and match:
Everything is logged. Every message, every tool call, every thought an agent writes to its scratchpad, so you can review a run afterward or live-tail it as it happens. The most valuable findings tend to be “that’s strange... why did it keep doing that?”, and agentspace is built to surface those moments and let you re-run them to test whether the effect is real.