Open-source core supported by a 2026 grant from the Foresight Institute. Independent project; ongoing.

You are viewing the web demo. This is a semi-functional demo.
Please see the fully functional command-line tool on GitHub at github.com/sfgeekgit/agentspace.

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 empirically. In practice it is a tool for spinning up many copies of an experiment in parallel — running the same situation over and over, changing one thing or nothing at all — and watching for anything interesting or surprising.

A World is a self-contained environment holding one or more agents — their memories, their messages, and the situation they are in.

Snapshot a running world at any moment, then fork the snapshot into a fresh copy. Change something big, or something as small as a single character, or change nothing at all — then run it again. Run twenty copies if you like. Run → snapshot → tweak → fork → compare is the whole loop. Each snapshot is a complete container image published to a public registry (ghcr.io), so anyone can pull a world down and reproduce or extend an experiment exactly — and no keys ever travel with it.

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

• Scenarios define a situation — a deception game, an open-ended task, an iterated prisoner’s dilemma, cooperation under scarcity. A core focus of agentspace is letting researchers create and edit their own scenarios.
• Modules add optional mechanics to most scenarios: enforceable contracts, prediction markets, bounded secret side-channels, the ability for agents to clone themselves, and more. Most modules combine freely with most scenarios.
• Personas set each agent’s starting character — its starting “soul.” Most souls can be dropped into any scenario or role. Each agent runs on any backend model and inside any runtime. Mix and match any combination.

Everything is logged — every message, every tool call, every thought an agent writes to its scratchpad. Researchers get easy access to the logs, whether reviewing them after the fact or live-tailing a run as it happens. With most scenarios and runtimes, you can also live-chat with the agents mid-run.

Spinning up copies or variants of a world is a single operation. agentspace handles the plumbing: securely minting a fresh API key per world, provisioning new DigitalOcean droplets on demand (or reusing existing ones, or running locally), storing and retrieving container images, and tearing it all down afterward. That leaves the researcher free to focus on designing scenarios and watching what happens.

This is exploratory infrastructure by design. The most valuable findings about agent behavior tend to be “that’s strange — why did it keep doing that?” rather than clean confirmations of a hypothesis. agentspace is built to surface those moments: worlds rich enough for surprising dynamics, with observability deep enough to see why a run cooperated, defected, or manipulated — then re-run it to test whether the effect is real.

github.com/sfgeekgit/agentspace ›