Agent-to-Agent Protocol Design: Building Interoperable Multi-Agent Communication

The Interoperability Problem in Multi-Agent Systems

When you build a single agent, communication is straightforward — the agent calls tools and returns results. When you build a team of agents, you face a fundamental question: how do agents talk to each other? Without a well-designed protocol, you end up with tightly coupled agents that can only work with the specific partners they were built for.

Agent-to-agent (A2A) protocol design solves this by establishing standard message formats, capability discovery, and negotiation patterns that allow any agent to communicate with any other agent — even agents built by different teams or frameworks.

Designing the Message Schema

Every inter-agent message needs a consistent structure. Here is a protocol envelope that supports request-response, streaming, and event patterns.

flowchart TD
    INPUT(["Task input"])
    SUPER["Supervisor agent<br/>plans plus monitors"]
    W1["Worker 1<br/>research"]
    W2["Worker 2<br/>code"]
    W3["Worker 3<br/>writing"]
    CRITIC{"Output meets<br/>rubric?"}
    REWORK["Rework or<br/>retry path"]
    SHARED[("Shared scratchpad<br/>and memory")]
    OUT(["Final result"])
    INPUT --> SUPER
    SUPER --> W1 --> CRITIC
    SUPER --> W2 --> CRITIC
    SUPER --> W3 --> CRITIC
    W1 --> SHARED
    W2 --> SHARED
    W3 --> SHARED
    SHARED --> SUPER
    CRITIC -->|Pass| OUT
    CRITIC -->|Fail| REWORK --> SUPER
    style SUPER fill:#4f46e5,stroke:#4338ca,color:#fff
    style CRITIC fill:#f59e0b,stroke:#d97706,color:#1f2937
    style OUT fill:#059669,stroke:#047857,color:#fff
    style SHARED fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b

from dataclasses import dataclass, field
from enum import Enum
from typing import Any, Dict, Optional
import uuid
import time

class MessageType(Enum):
    REQUEST = "request"
    RESPONSE = "response"
    EVENT = "event"
    CAPABILITY_QUERY = "capability_query"
    CAPABILITY_ADVERTISEMENT = "capability_advertisement"
    NEGOTIATION = "negotiation"

@dataclass
class AgentMessage:
    msg_type: MessageType
    sender_id: str
    receiver_id: str
    payload: Dict[str, Any]
    correlation_id: str = field(default_factory=lambda: str(uuid.uuid4()))
    reply_to: Optional[str] = None
    timestamp: float = field(default_factory=time.time)
    ttl_seconds: float = 30.0
    protocol_version: str = "1.0"

    @property
    def is_expired(self) -> bool:
        return (time.time() - self.timestamp) > self.ttl_seconds

    def create_reply(self, payload: Dict[str, Any]) -> "AgentMessage":
        return AgentMessage(
            msg_type=MessageType.RESPONSE,
            sender_id=self.receiver_id,
            receiver_id=self.sender_id,
            payload=payload,
            reply_to=self.correlation_id,
        )

The correlation_id ties requests to responses. The reply_to field enables threading. The ttl_seconds prevents stale messages from clogging the system.

Capability Advertisement and Discovery

Agents need to discover what other agents can do. A capability registry lets agents advertise their skills and query for agents that match specific needs.

from dataclasses import dataclass
from typing import Dict, List, Set

@dataclass
class AgentCapability:
    name: str
    description: str
    input_schema: Dict     # JSON Schema for expected input
    output_schema: Dict    # JSON Schema for output
    cost_estimate: float   # Relative cost (0.0 to 1.0)
    latency_ms: int        # Expected latency

class CapabilityRegistry:
    def __init__(self):
        self._capabilities: Dict[str, List[AgentCapability]] = {}

    def register(self, agent_id: str, capabilities: List[AgentCapability]):
        self._capabilities[agent_id] = capabilities

    def unregister(self, agent_id: str):
        self._capabilities.pop(agent_id, None)

    def find_agents(
        self, capability_name: str, max_cost: float = 1.0
    ) -> List[str]:
        results = []
        for agent_id, caps in self._capabilities.items():
            for cap in caps:
                if (
                    cap.name == capability_name
                    and cap.cost_estimate <= max_cost
                ):
                    results.append(agent_id)
        return results

    def get_capabilities(self, agent_id: str) -> List[AgentCapability]:
        return self._capabilities.get(agent_id, [])

Negotiation Protocol

When multiple agents can handle a task, the requesting agent negotiates to find the best match. This implements a simple contract-net protocol.

import asyncio

class NegotiationProtocol:
    def __init__(self, registry: CapabilityRegistry, bus: "MessageBus"):
        self.registry = registry
        self.bus = bus

    async def request_bids(
        self,
        requester_id: str,
        capability_name: str,
        task_payload: Dict,
        timeout: float = 5.0,
    ) -> List[Dict]:
        candidates = self.registry.find_agents(capability_name)
        if not candidates:
            return []

        bids = []
        bid_requests = []
        for agent_id in candidates:
            msg = AgentMessage(
                msg_type=MessageType.NEGOTIATION,
                sender_id=requester_id,
                receiver_id=agent_id,
                payload={
                    "action": "request_bid",
                    "capability": capability_name,
                    "task": task_payload,
                },
            )
            bid_requests.append(self.bus.send_and_wait(msg, timeout))

        results = await asyncio.gather(
            *bid_requests, return_exceptions=True
        )
        for result in results:
            if isinstance(result, AgentMessage):
                bids.append(result.payload)
        return bids

    def select_winner(self, bids: List[Dict]) -> Optional[Dict]:
        valid = [b for b in bids if b.get("accepted")]
        if not valid:
            return None
        return min(valid, key=lambda b: b.get("cost", float("inf")))

Message Bus Implementation

The message bus routes messages between agents and supports both direct addressing and publish-subscribe patterns.

class MessageBus:
    def __init__(self):
        self._handlers: Dict[str, asyncio.Queue] = {}

    def register_agent(self, agent_id: str):
        self._handlers[agent_id] = asyncio.Queue()

    async def send(self, message: AgentMessage):
        queue = self._handlers.get(message.receiver_id)
        if queue:
            await queue.put(message)

    async def send_and_wait(
        self, message: AgentMessage, timeout: float = 10.0
    ) -> Optional[AgentMessage]:
        await self.send(message)
        queue = self._handlers.get(message.sender_id)
        if not queue:
            return None
        try:
            while True:
                reply = await asyncio.wait_for(queue.get(), timeout)
                if reply.reply_to == message.correlation_id:
                    return reply
        except asyncio.TimeoutError:
            return None

    async def receive(
        self, agent_id: str, timeout: float = 5.0
    ) -> Optional[AgentMessage]:
        queue = self._handlers.get(agent_id)
        if not queue:
            return None
        try:
            return await asyncio.wait_for(queue.get(), timeout)
        except asyncio.TimeoutError:
            return None

Putting It Together

async def demo():
    bus = MessageBus()
    registry = CapabilityRegistry()

    # Register agents and their capabilities
    bus.register_agent("summarizer")
    registry.register("summarizer", [
        AgentCapability("summarize", "Summarize text", {}, {}, 0.3, 2000)
    ])

    bus.register_agent("translator")
    registry.register("translator", [
        AgentCapability("translate", "Translate text", {}, {}, 0.5, 3000)
    ])

    # Discover and negotiate
    negotiator = NegotiationProtocol(registry, bus)
    agents = registry.find_agents("summarize")
    print(f"Agents with summarize capability: {agents}")

FAQ

Why not just use HTTP REST between agents?

HTTP REST works for simple request-response patterns but lacks built-in support for capability discovery, negotiation, and message correlation. A dedicated agent protocol gives you these features plus TTL-based expiration and structured negotiation — reducing the boilerplate each agent must implement.

How does this compare to the A2A protocol from Google?

Google's Agent-to-Agent protocol focuses on web-standard interoperability using JSON-RPC over HTTP with agent cards for discovery. The patterns in this article follow similar principles but are designed for in-process or single-cluster deployments. For cross-organization interoperability, adopt the A2A standard; for internal agent teams, a lightweight custom protocol often performs better.

---

#AgentProtocol #A2ACommunication #MultiAgentSystems #AgentInteroperability #ProtocolDesign #AgenticAI #PythonAI #DistributedAgents