Agent Protocol

Overview

Agent Protocol answers a simple question: if you want to deploy an AI agent as a service, what should the API look like?

It provides an OpenAPI specification that any agent—regardless of framework—can implement to expose a consistent REST interface.

Core Endpoints

POST /ap/v1/agent/tasks           # Create a new task
GET  /ap/v1/agent/tasks           # List all tasks
GET  /ap/v1/agent/tasks/{id}      # Get task details
POST /ap/v1/agent/tasks/{id}/steps  # Execute next step
GET  /ap/v1/agent/tasks/{id}/steps  # List task steps

Key Features

No SDK Required

Agent Protocol uses standard HTTP—you can interact with agents using cURL, Postman, or any HTTP client.

Framework Agnostic

Implement the API in any language/framework. AutoGPT, LangChain agents, and custom implementations all work.

Step-Based Execution

Tasks are executed step-by-step, allowing for:

• Progress tracking
• Intermediate results
• Human intervention between steps

W3C Community Group

The W3C has established an AI Agent Protocol Community Group to:

• Develop open, interoperable protocols
• Enable agent discovery across the web
• Standardize agent communication

Specification: https://w3c-cg.github.io/ai-agent-protocol/protocol.html

Relationship with A2A

Agent Protocol and A2A serve different purposes:

Aspect	Agent Protocol	A2A
Focus	HTTP API surface	Agent communication
Use Case	Exposing single agent	Multi-agent collaboration
Complexity	Simple REST	Rich task lifecycle

Some projects are discussing migration to A2A for agent-to-agent scenarios while keeping Agent Protocol for simple API exposure.

When to Use Agent Protocol

Use Agent Protocol for:

• Deploying agents as REST services
• Simple agent integrations
• AutoGPT-compatible implementations

Consider A2A for:

• Multi-agent systems
• Rich task negotiation
• Enterprise deployments

# Create and execute an agent task
import requests

# Create a task
response = requests.post(
    "https://agent.example.com/ap/v1/agent/tasks",
    json={"input": "Summarize this document..."}
)
task_id = response.json()["task_id"]

# Execute steps until complete
while True:
    step = requests.post(
        f"https://agent.example.com/ap/v1/agent/tasks/{task_id}/steps"
    ).json()

    if step["is_last"]:
        print(f"Result: {step['output']}")
        break