Agent Mesh Protocol

Protocol Overview

The Agent Mesh Protocol (AMP) is a standardized communication protocol designed specifically for autonomous AI agents operating in distributed environments. AMP enables agents built with different frameworks to discover, communicate, and collaborate seamlessly.

Key Design Principles

• Framework Agnostic: Works with any AI agent framework
• Transport Independent: Supports HTTP, WebSocket, gRPC, and message queues
• Capability-Based: Agents advertise and discover capabilities dynamically
• Security First: Built-in authentication, authorization, and encryption
• Backward Compatible: Semantic versioning ensures smooth upgrades

Message Format

All AMP messages follow a standardized JSON format with required metadata and flexible payload structure.

{
  "protocol": "AMP/1.0",
  "message": {
    "id": "unique-message-id",
    "type": "request|response|event|error",
    "timestamp": "2025-01-27T10:30:00Z",
    "source": {
      "agent_id": "source-agent-id",
      "agent_name": "Source Agent",
      "framework": "langchain"
    },
    "destination": {
      "agent_id": "target-agent-id",
      "capability": "text-analysis"
    },
    "headers": {
      "correlation_id": "request-correlation-id",
      "priority": 5,
      "timeout_ms": 30000,
      "retry_count": 0
    },
    "payload": {
      // Capability-specific data
    }
  }
}

Required Fields

• protocol - Protocol version
• message.id - Unique message ID
• message.type - Message type
• message.timestamp - ISO 8601 timestamp
• message.source - Source agent info

Optional Fields

• message.destination - Target agent/capability
• message.headers - Additional metadata
• message.payload - Message content

Message Types

Request

Initiated by an agent to invoke a capability on another agent.

{
  "type": "request",
  "destination": {
    "agent_id": "analyzer-001",
    "capability": "sentiment-analysis"
  },
  "payload": {
    "text": "This product is amazing!",
    "options": {
      "confidence_threshold": 0.8
    }
  }
}

Response

Sent by an agent in reply to a request message.

{
  "type": "response",
  "headers": {
    "correlation_id": "req-123-456"
  },
  "payload": {
    "result": {
      "sentiment": "positive",
      "confidence": 0.92,
      "categories": ["satisfaction", "enthusiasm"]
    },
    "status": "success"
  }
}

Event

Broadcast by agents to notify others of state changes or events.

{
  "type": "event",
  "payload": {
    "event_type": "capability_registered",
    "capability": "image-classification",
    "agent_id": "vision-agent-002",
    "metadata": {
      "model": "resnet-50",
      "supported_formats": ["jpg", "png", "webp"]
    }
  }
}

Error

Indicates that an error occurred during message processing.

{
  "type": "error",
  "headers": {
    "correlation_id": "req-123-456"
  },
  "payload": {
    "error_code": "CAPABILITY_NOT_FOUND",
    "error_message": "Requested capability 'advanced-reasoning' not available",
    "retry_after": 300,
    "suggested_alternatives": ["basic-reasoning", "rule-based-inference"]
  }
}

Capability System

AMP uses a hierarchical capability taxonomy to enable agents to discover and invoke functionality across the mesh network.

Capability Declaration

{
  "agent_id": "nlp-processor-001",
  "capabilities": [
    {
      "id": "text-analysis",
      "version": "1.2.0",
      "description": "Advanced text analysis and processing",
      "input_schema": {
        "type": "object",
        "properties": {
          "text": {"type": "string", "maxLength": 10000},
          "language": {"type": "string", "default": "en"}
        },
        "required": ["text"]
      },
      "output_schema": {
        "type": "object",
        "properties": {
          "tokens": {"type": "array"},
          "entities": {"type": "array"},
          "sentiment": {"type": "string"}
        }
      },
      "constraints": {
        "max_response_time_ms": 5000,
        "rate_limit": "100/minute",
        "supported_languages": ["en", "es", "fr", "de"]
      }
    }
  ]
}

Text Processing

• text-analysis
• text-summarization
• text-translation
• text-classification

Generation

• gen-text
• gen-code
• gen-structured-data
• gen-creative-content

Reasoning

• qa-factual
• qa-reasoning
• analysis-pattern
• plan-task-decomposition

View complete capability taxonomy

Transport Bindings

AMP is transport-agnostic and can be implemented over various communication protocols.

HTTP/REST

Synchronous request-response pattern over HTTP.

POST /amp/v1/invoke
Content-Type: application/json

{
  "protocol": "AMP/1.0",
  "message": { ... }
}

WebSocket

Bidirectional, persistent connections for real-time communication.

ws://agent-mesh.example.com/amp/v1
  
// Send AMP message as JSON
ws.send(JSON.stringify(ampMessage))

Message Queues

Asynchronous messaging via AMQP, Kafka, or similar systems.

Topic: amp.requests.text-analysis
Routing Key: agent.nlp-processor-001

Message: AMP JSON payload

gRPC

High-performance RPC with Protocol Buffers serialization.

service AMPService {
  rpc InvokeCapability(AMPMessage) 
    returns (AMPMessage);
}

Security

AMP implements multiple layers of security to ensure safe agent communication.

Authentication

Supported Methods

• API Key authentication
• JWT tokens
• mTLS certificates
• HMAC signing

Authorization

• Capability-based access control
• Role-based permissions
• Agent identity verification
• Resource quotas and limits

Message Integrity

All messages can be cryptographically signed to ensure authenticity and prevent tampering.

{
  "protocol": "AMP/1.0",
  "message": { ... },
  "signature": {
    "algorithm": "HS256",
    "keyId": "agent-key-001",
    "signature": "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9..."
  }
}

Protocol Specification