Skip to content

L13 - Transport & Flow Semantics

L13 covers the mechanics of AI calls as flows: streaming, cancellation, resume tokens, multiplexing, idempotency, retries, backpressure, and transport negotiation. It is where AI interaction becomes a durable protocol conversation rather than a single request-response call.

L13Transport & Flow Semantics
  1. Streaming
  2. Cancel and resume
  3. Backpressure
  4. Multiplexing

What belongs here

L13 is not simply "HTTP." It describes flow behavior that AI systems increasingly need: partial results, tool progress, long-running actions, resumable conversations, and reliable cancellation across provider and host boundaries.

Representative projects and protocols

Project or protocol Why it might fit Adjacent layers
MCP transports Transport choices and message flow for Model Context Protocol clients and servers. L10 tools, L13 transport
A2A protocol Agent-to-agent protocol work that raises flow, task, and interaction semantics. L13 transport, L12 planning
gRPC RPC framework with streaming patterns useful for service-to-service AI systems. L13 transport, L16 services
NATS Messaging system for distributed systems and event-driven coordination. L13 transport, L12 orchestration
Server-Sent Events Common browser-facing streaming primitive used by AI APIs and apps. L13 streaming, L16 UX
Temporal Durable execution framework relevant to long-running AI workflows. L13 durability, L12 planning
MCP-AQL adapters and reference implementations Public reference runtime work where streaming behavior, execution lifecycle, and transport expectations become concrete. L13 transport, L10 invocation

Boundary questions

  • What semantics are AI-specific, and what should remain generic distributed-systems plumbing?
  • Should resumability be modeled in L13 transport, L14 session, or both?
  • How should tool progress, model streaming, and human approval pauses share one flow vocabulary?

Signals to watch

  • More AI protocols adding cancellation, resume, and progress semantics.
  • Browser and local-app AI clients needing richer streaming and session recovery.
  • Durable agent workflows making "request-response" insufficient as the main mental model.