By Sagar Shankaran, Founder of CallSphere
Implement event-driven communication between AI agent microservices using Kafka, RabbitMQ, and NATS. Learn event schema design, pub/sub patterns, event sourcing, and exactly-once delivery semantics.
Key takeaways
AI agent workflows are inherently asynchronous. A user sends a message, the agent reasons over it, calls tools, retrieves context from a vector store, and eventually returns a response. Many of these steps can happen independently. The memory service needs to record the conversation after the response is sent. The analytics service needs to log latency metrics. The billing service needs to track token usage.
If all of these happen synchronously in the request path, response latency balloons. Event-driven architecture decouples the request path from downstream processing. The conversation service publishes events, and other services consume them independently.
A well-designed event schema is the contract between services. It must be self-describing, versioned, and contain enough context for any consumer to act without making additional API calls:
flowchart LR
INPUT(["User intent"])
PARSE["Parse plus<br/>classify"]
PLAN["Plan and tool<br/>selection"]
AGENT["Agent loop<br/>LLM plus tools"]
GUARD{"Guardrails<br/>and policy"}
EXEC["Execute and<br/>verify result"]
OBS[("Trace and metrics")]
OUT(["Outcome plus<br/>next action"])
INPUT --> PARSE --> PLAN --> AGENT --> GUARD
GUARD -->|Pass| EXEC --> OUT
GUARD -->|Fail| AGENT
AGENT --> OBS
style AGENT fill:#4f46e5,stroke:#4338ca,color:#fff
style GUARD fill:#f59e0b,stroke:#d97706,color:#1f2937
style OBS fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b
style OUT fill:#059669,stroke:#047857,color:#fff
from dataclasses import dataclass, field, asdict
from datetime import datetime
import uuid
import json
@dataclass
class AgentEvent:
event_id: str = field(default_factory=lambda: str(uuid.uuid4()))
event_type: str = ""
version: str = "1.0"
timestamp: str = field(
default_factory=lambda: datetime.utcnow().isoformat()
)
source_service: str = ""
correlation_id: str = ""
payload: dict = field(default_factory=dict)
def to_json(self) -> str:
return json.dumps(asdict(self))
# Example events published by the conversation service
def create_message_received_event(
session_id: str, user_msg: str, correlation_id: str
) -> AgentEvent:
return AgentEvent(
event_type="agent.message.received",
source_service="conversation-manager",
correlation_id=correlation_id,
payload={
"session_id": session_id,
"message": user_msg,
"message_type": "user",
},
)
def create_response_generated_event(
session_id: str,
response: str,
tokens_used: int,
model: str,
correlation_id: str,
) -> AgentEvent:
return AgentEvent(
event_type="agent.response.generated",
source_service="conversation-manager",
correlation_id=correlation_id,
payload={
"session_id": session_id,
"response_length": len(response),
"tokens_used": tokens_used,
"model": model,
},
)
The correlation_id ties all events from a single user request together across services, which is essential for distributed tracing.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
Kafka excels when you need durable, ordered event streams at high throughput. Agent systems that process thousands of messages per minute benefit from Kafka's partitioned log architecture:
from aiokafka import AIOKafkaProducer, AIOKafkaConsumer
import asyncio
# Producer in the conversation service
class AgentEventProducer:
def __init__(self, bootstrap_servers: str = "kafka:9092"):
self.producer = AIOKafkaProducer(
bootstrap_servers=bootstrap_servers,
value_serializer=lambda v: v.encode("utf-8"),
acks="all", # Wait for all replicas to acknowledge
)
async def start(self):
await self.producer.start()
async def publish(self, event: AgentEvent):
topic = event.event_type.replace(".", "-")
await self.producer.send_and_wait(
topic=topic,
value=event.to_json(),
key=event.correlation_id.encode("utf-8"),
)
# Consumer in the analytics service
class AnalyticsConsumer:
def __init__(self):
self.consumer = AIOKafkaConsumer(
"agent-response-generated",
bootstrap_servers="kafka:9092",
group_id="analytics-service",
auto_offset_reset="earliest",
enable_auto_commit=False,
)
async def consume(self):
await self.consumer.start()
try:
async for msg in self.consumer:
event = json.loads(msg.value.decode("utf-8"))
await self.process_event(event)
await self.consumer.commit()
finally:
await self.consumer.stop()
async def process_event(self, event: dict):
payload = event["payload"]
await self.db.insert_metric(
session_id=payload["session_id"],
tokens_used=payload["tokens_used"],
model=payload["model"],
timestamp=event["timestamp"],
)
Setting acks="all" ensures the event is durably written before the producer considers it sent. The consumer uses manual commit (enable_auto_commit=False) to guarantee at-least-once processing.
NATS is a strong choice for agent systems that need low-latency pub/sub without Kafka's operational complexity:
import nats
async def nats_publisher():
nc = await nats.connect("nats://nats:4222")
event = create_message_received_event(
session_id="sess-123",
user_msg="What is my account balance?",
correlation_id="req-abc",
)
await nc.publish(
"agent.message.received",
event.to_json().encode(),
)
await nc.flush()
await nc.close()
async def nats_subscriber():
nc = await nats.connect("nats://nats:4222")
sub = await nc.subscribe("agent.>") # Wildcard subscription
async for msg in sub.messages:
event = json.loads(msg.data.decode())
print(f"Received {event['event_type']} "
f"from {event['source_service']}")
NATS uses subject-based addressing with wildcards. The pattern agent.> subscribes to all events under the agent namespace, making it easy to build monitoring dashboards.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
True exactly-once delivery is achievable through idempotent consumers. Store the event_id in a processed-events table and check it before processing:
async def process_event_exactly_once(self, event: dict):
event_id = event["event_id"]
if await self.db.event_already_processed(event_id):
return # Skip duplicate
await self.handle(event)
await self.db.mark_event_processed(event_id)
Choose Kafka when you need durable event storage for replay, strict ordering within partitions, and high throughput at scale (thousands of events per second). Choose NATS when you need simple pub/sub with low latency, the event volume is moderate, and you want minimal operational overhead. For most agent systems under 500 requests per minute, NATS is simpler to operate.
Include a version field in every event. When the schema changes, increment the version. Consumers should handle multiple versions by checking the version field and applying the appropriate deserialization logic. Avoid breaking changes — add new fields rather than renaming or removing existing ones.
Every service that performs a meaningful state change should publish events. The tool execution service should publish tool.execution.completed events. The RAG service should publish rag.retrieval.completed events. This gives downstream services full visibility into the agent's behavior without coupling them to the conversation service.
#EventDriven #Kafka #RabbitMQ #NATS #Microservices #AgenticAI #LearnAI #AIEngineering

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
The 2026 desktop AI agent landscape — ServiceNow Project Arc, Anthropic Claude offerings, OpenAI agents, and Google Mariner. A buyer's map.
An agentic-AI perspective on Anthropic Skills system, covering orchestration patterns, tool use, and how agent tooling fits production agent stacks.
Enterprise CIO Guide perspective on Comet's general-availability launch put an agentic browser in front of millions of consumers, and it works better than the demos suggested.
Enterprise CIO Guide perspective on Harvey AI's enterprise rollout numbers show legal agents have moved past the pilot stage at AmLaw 100 firms.
Enterprise CIO Guide perspective on Hippocratic AI's deployment numbers show healthcare voice agents are moving from pilot to production across major US health systems.
An agentic-AI perspective on Claude Agent SDK loops, covering orchestration patterns, tool use, and how agent orchestration fits production agent stacks.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.