ARBITER

AI agents are beginning to communicate directly, on behalf of humans: in procurement, data sharing, service contracting, and beyond. The interactions between them need to be more than functional. They need to be trustworthy.

The risks of AI systems being manipulated are well documented in human-agent settings. Agent-to-agent interaction creates a new and largely unaddressed attack surface. Two goal-directed systems communicating without a human in the loop can produce behaviours that neither agents' principals anticipated or intended.

ARBITER is an open-source trusted interaction monitor for agent-to-agent negotiation. It watches agent interactions, detects adversarial, deceptive and manipulative behaviour, and produces a cryptographic guarantees of the honesty and fairness of the interaction, usable by both parties and downstream dispute resolvers.

Humans manage equivalent problems through social norms, reputation, and neutral arbiters: courts, mediators, regulators. ARBITER provides the same assurance for autonomous agents.