Toward a network for AI agents

Let the workload define the network question

We are taking PqP’s next research phase toward communication between independently operated AI agents. A coding agent on one machine may hand work to another runtime, go offline, and return after the result is ready. The communication layer has to account for that whole interval.

PqP’s foundation combines DNS-based discovery, encrypted queue-assisted messaging, and direct WebRTC transport. The agent workload gives us a concrete way to investigate when those paths help: participants may have no permanent process, public inbound endpoint, or shared operator.

A task needs more than a message

The proposed first interoperability target is A2A. An adapter would carry its task semantics while PqP handles contact and delivery. This integration is planned; the current prototype does not yet establish A2A compatibility.

The receiver needs to know who requested the task, what they are permitted to ask, how progress is represented, and what happens after a retry. Network reachability cannot answer those questions on its own. Owner authority and task state have to remain explicit throughout an exchange.

Measure completion through interruption

The study will use bounded tasks and explicit failure cases. We want to distinguish a connection, an accepted message, and a completed task, then compare the approach with simpler transports under the same conditions.

• Identity and authority: establish the sender and the scope of the requested action.
• Delivery and recovery: account for duplicate messages, expiry, retries, and restarts.
• Path changes: preserve task state when moving between queued and direct communication.
• Cost and benefit: measure latency, bandwidth, infrastructure use, and task completion against conventional alternatives.

The work ahead

Identity and control-plane security remain the live engineering priority, followed by delivery semantics. Agent interoperability builds on those gates. Public agent discovery, cross-owner authorization, and paid agent services are not available capabilities of the prototype.

The next research output should make the workload, threat model, and comparison reproducible. The purpose of this direction is to establish where queue-assisted peer communication earns its complexity in an agent system.