Vision

AEP Vision

One-Line Vision

AEP makes asynchronous communication a first-class capability for LLM agents.

Problem

Modern agent systems are no longer simple request-response programs. They run background tasks, watch external environments, update memory, collaborate with other agents, receive delayed tool results, and react to context changes after the original prompt has moved on.

MCP standardizes how agents synchronously discover and call tools. That is necessary but incomplete. Agents also need a standard way to receive events, progress, state changes, memory updates, and deferred results.

Without such a protocol, every runtime invents its own pattern:

These approaches do not compose well across agent runtimes, tools, memory systems, and deployment environments.

Vision

AEP should become the common asynchronous event layer for agentic systems.

An agent should be able to subscribe to a memory stream, launch a long-running tool task, receive progress events, observe external state changes, cancel work, replay missed events, and correlate all of this with a conversation or task without depending on one vendor's runtime.

Core Use Cases

Event Streams

Agents need to receive events from tools, environments, memory systems, orchestrators, and other agents.

Examples:

Async Tool Calls

Some tool calls are too slow or open-ended for synchronous RPC.

Examples:

AEP should represent these as tasks with accepted, progress, output, completed, failed, cancelled, and timed-out events.

Context And Memory Events

Agents need to react when their context changes.

Examples:

Agent-To-Agent Communication

Agents need to exchange messages without blocking each other.

Examples:

Goals

Non-Goals

Design Principles