SWARM
A research framework for measuring emergent safety failures that appear only when many AI agents interact, even when each individual agent behaves safely on its own.
SWARM, standing for System-Wide Assessment of Risk in Multi-agent systems, is a research framework for measuring emergent failures that appear only at the level of interaction between many AI agents, not at the level of any single agent. It provides interaction-level safety metrics such as an illusion-delta measure of replay variability, and governance experiments covering audits, staking and sanctions.
A minimal example runs a three-agent simulation with one deceptive actor and computes an illusion-delta signal from replay variability in about sixty seconds. The project frames its motivation around information asymmetry, adverse selection and capability asymmetry between agents, and between agents and the humans who delegate to them, arguing these risks do not require any single agent to reach AGI-level capability. It is written in Python and MIT licensed.