The agent ecosystem is experiencing a fundamental architectural fork in communication protocols. On one side, unidirectional streaming—used by Claude Code and similar systems—spawns processes and streams JSONL via stdout. On the other, bidirectional RPC—exemplified by Hermes and the ACP protocol—maintains persistent JSON-RPC servers with active notifications from agents. This article argues that the fork should happen at the transport layer, not the parsing layer, and provides a detailed breakdown of the ACP protocol's four key components: message framing, request-response matching, notification handling, and error propagation. For developers building multi-agent systems, this choice directly impacts latency, reliability, and the ability to handle complex workflows. The ACP approach offers stronger guarantees for production systems, while unidirectional streaming remains simpler for prototyping. Understanding this trade-off is essential for anyone designing agent infrastructure.
Deep technical analysis of the two dominant agent communication paradigms and why transport-layer separation is critical.