Agent Communication Protocol

The Agent Communication Protocol (ACP) is an open standard developed by IBM Research for enabling seamless, interoperable communication between AI agents across frameworks, languages, and organizations via a unified RESTful HTTP interface. Introduced through the BeeAI platform, ACP addresses ecosystem fragmentation by standardizing agent discovery, task delegation, and messaging without dictating internal agent implementations.

Overview

ACP operates on a client-server model where clients (agents, humans, or services) send HTTP requests to server endpoints wrapping AI agents. Unlike JSON-RPC-based protocols, ACP uses standard REST conventions with JSON payloads, making it accessible via curl, Postman, or any HTTP client without requiring specialized SDKs.

The protocol defines structured message envelopes containing task IDs, metadata, content payloads, and optional streaming channels. Agents publish lightweight manifests describing their capabilities, schemas, and status for both online and offline discovery, supporting scale-to-zero deployments in cloud-native environments.

Core Architecture

ACP's architecture centers on several key mechanisms:

HTTP-Based Messaging — Standard REST endpoints (POST for messages) with JSON payloads for routing and task resumption
Asynchronous Support — Default async mode via fire-and-forget POSTs with task IDs; clients poll or subscribe via SSE/WebSockets for progress deltas or completion
Synchronous Support — Simple HTTP POST/JSON responses for low-latency use cases
Agent Discovery — Manifest-based capability advertisement for runtime and offline lookup
Lifecycle Management — Defined agent states (INITIALIZING, ACTIVE, RETIRED) with session IDs for stateful interactions
Multimodal Content — Native support for text, images, audio, video, and binary data via MIME types

Comparison with MCP and A2A

ACP exists alongside two other major agent protocols: Anthropic's Model Context Protocol (MCP) and Google's Agent-to-Agent (A2A) protocol. Each addresses different communication layers:

Aspect	IBM ACP	Anthropic MCP	Google A2A
Core Protocol	REST/HTTP	JSON-RPC 2.0	JSON-RPC
Primary Focus	Agent-to-agent ops	Tool and data access	Agent collaboration
Async/Streaming	Default async + SSE/WebSockets	Task state polling	Collaborative tasks
Multimodal	Full MIME support	Primarily tools/data	Metadata-focused
Discovery	Online/offline manifests	Framework-tied	Standardized metadata
SDK Required	Optional	Often required	Varies

ACP is intentionally complementary: MCP connects agents to tools and data sources, A2A handles multi-agent collaboration patterns, and ACP provides the lightweight HTTP transport layer for agent-to-agent messaging.

Code Example

Sending a message to an ACP-compliant agent:

import requests
 
ACP_ENDPOINT = "https://agent-server.example.com/api/v1/messages"
AGENT_ID = "weather-agent-01"
 
payload = {
    "task_id": "task-abc-123",
    "agent_id": AGENT_ID,
    "content": {
        "type": "text/plain",
        "body": "What is the weather forecast for Tokyo?"
    },
    "metadata": {
        "session_id": "session-xyz",
        "priority": "normal"
    }
}
 
response = requests.post(ACP_ENDPOINT, json=payload)
result = response.json()
print(f"Status: {result['status']}, Response: {result['content']['body']}")

Discovering available agents via manifest:

import requests
 
DISCOVERY_URL = "https://agent-server.example.com/api/v1/agents"
 
agents = requests.get(DISCOVERY_URL).json()
for agent in agents["agents"]:
    print(f"Agent: {agent['name']}, Capabilities: {agent['capabilities']}")
    print(f"  Status: {agent['lifecycle_state']}")

Security

ACP incorporates several security mechanisms:

Capability Tokens — Signed, unforgeable tokens with expiry for authorization
OTLP Observability — OpenTelemetry integration for monitoring and tracing
Kubernetes RBAC — Native integration with container orchestration access control
Transport Security — Standard HTTPS/TLS for all communications

Table of Contents