When two agents cannot communicate directly — one is behind a firewall, the other is intermittently connected — a relay accepts messages from the sender, stores them, and delivers them when the recipient is reachable. This is the same pattern as email’s SMTP relay infrastructure, adapted for agent-to-agent communication.
What distinguishes an IFP relay from a generic message broker is mandatory accountability. Every relay must append a trace entry to every message it handles (IFP-8, implementing IFP-1 Section 6.4). The trace records when the relay received the message, from whom, and any transformations applied. This trace is visible to the receiving agent and, through the audit log, to the human operator.
Relays exist on a trust spectrum. An untrusted relay (the default assumption) may log message content — agents should encrypt bodies when routing through untrusted relays. A trusted relay has a known operator. A self-operated relay is run by one of the communicating parties.
The relay model preserves two properties that matter for human agency: signatures survive relay (IFP-5 signing covers the message, not the transport — a relay cannot forge the sender’s signature), and trace cannot be hidden (mandatory trace entries make relay involvement visible to humans reviewing the audit log).