Every few weeks someone asks me how their agents should talk to each other. The answer I give depends on what they’re actually trying to do, and I’ve noticed most people pick a method based on what they already know rather than what fits the problem. So I sat down and actually tested all six approaches I’ve seen in production.
Same test each time: Agent A sends a request, Agent B processes it and responds, I measure round-trip time and what breaks under real network conditions. Both agents running on separate machines, one of them behind NAT.
Here’s what I found.
1. HTTP Polling
The oldest pattern and still the most common. Agent A calls Agent B’s REST endpoint on an interval and checks for updates.
import httpx, time
while True:
response = httpx.get("http://agent-b/status")
if response.json()["ready"]:
result = httpx.get("http://agent-b/result")
break
time.sleep(2)
Round-trip when the result is ready: depends entirely on your poll interval. With 2-second polling, average wait is 1 second even if Agent B responds in 50ms. With 500ms polling you’re burning requests 120 times per minute per agent pair.
The HTTP/1.1 spec was never designed for this use case and it shows. Connection overhead adds up fast at scale. You also need Agent B to have a stable, reachable address — which is fine in a controlled cloud environment and completely breaks the moment either agent is behind NAT.
I wrote more about why I stopped using this for persistent agent connections here.
Verdict: Works. Wasteful. Fine for low-frequency checks, bad for anything real-time.
2. Webhooks
Agent B calls Agent A when it’s ready, instead of waiting to be asked. Lower latency than polling, simpler than keeping a connection open.
from flask import Flask, request
app = Flask(__name__)
@app.route("/webhook", methods=["POST"])
def receive():
data = request.json
# handle the result
return "ok", 200
Latency when it works: 50 to 200ms depending on network conditions. That’s actually decent.
The problem is “when it works.” Webhooks require Agent A to have a publicly reachable endpoint. In production cloud deployments that’s fine. In development it means ngrok or equivalent. Behind carrier-grade NAT it doesn’t work at all. I’ve spent more time debugging webhook delivery failures than I care to admit — retries, timeouts, signature verification breaking after a load balancer rotation.
I replaced webhooks with persistent tunnels in one pipeline and wrote about what changed here. The short version: fewer moving parts, nothing breaks when IPs change.
Verdict: Good latency, terrible NAT story, operational overhead accumulates.
3. WebSockets
RFC 6455 defined WebSockets in 2011. A persistent full-duplex connection over a single TCP socket. Both sides can push messages without waiting to be asked.
import asyncio, websockets
async def agent_b():
async with websockets.serve(handle, "0.0.0.0", 8765):
await asyncio.Future()
async def handle(ws):
request = await ws.recv()
result = process(request)
await ws.send(result)
Round-trip in my tests: 10 to 40ms. That’s good. Connection setup is a one-time cost, after which messages are cheap. Works well when you have a stable server that both agents can reach.
The catch is that you need that server. WebSockets don’t help you connect two agents that are both behind NAT — you still need a relay or a broker that both can reach. Managing connection state, reconnects, and backpressure is also on you. For one-to-one agent pairs it’s fine. For dynamic agent topologies where agents come and go, it gets complicated fast.
Verdict: Solid choice for persistent one-to-one connections where you control both endpoints.
4. MQTT
MQTT 5.0 (the current spec, maintained by OASIS) is a publish-subscribe protocol built for low-bandwidth, high-latency environments. Originally designed for IoT sensors, it’s found a second life in agent messaging because the broker model maps naturally to fan-out scenarios.
import paho.mqtt.client as mqtt
client = mqtt.Client()
client.connect("broker", 1883)
client.subscribe("agent-b/results")
client.publish("agent-b/requests", payload=my_request)
Latency with a local broker: 5 to 20ms. With a cloud broker: 20 to 100ms depending on geography. The broker handles persistence, QoS levels, and fan-out — which is genuinely useful if you have multiple agents listening for the same events.
The broker is also the central point of failure. If it goes down, nothing communicates. You can cluster brokers for resilience but then you’re running infrastructure. I removed the message broker from one pipeline specifically because of this — the write-up is here. Also worth noting: MQTT is pub/sub, not request/response. Implementing request/response on top of it requires correlation IDs and reply topics, which works but adds boilerplate.
Verdict: Genuinely good for event fan-out. Overkill for direct agent-to-agent calls, and the broker dependency matters.
5. gRPC
gRPC runs on HTTP/2, uses Protocol Buffers for serialization, and supports bidirectional streaming. It’s what you reach for when you need fast, typed, synchronous calls between services.
service AgentB {
rpc Process(Request) returns (Response);
rpc Stream(Request) returns (stream Response);
}
Round-trip in my tests: 5 to 15ms. Fastest synchronous option on this list. Binary serialization means smaller payloads than JSON. The streaming support is legitimately useful for agents that need to send partial results.
The friction is the schema management. Every interface change means updating the .proto file and regenerating clients. For stable, well-defined agent interfaces this is fine — it enforces contracts. For agents that are still evolving, it slows things down. Same NAT problem as everything else: you need both agents to reach each other, which means you need infrastructure or a proxy.
I’ve written about choosing between messaging protocols for agent-to-agent communication here if you want more on the gRPC vs alternatives tradeoffs.
Verdict: Best raw performance for synchronous calls. Schema overhead is real. NAT is still your problem.
6. Pilot Protocol
Pilot Protocol is a peer-to-peer overlay network built specifically for agents. Each node gets a virtual address derived from its Ed25519 keypair. NAT traversal uses STUN with hole-punching and relay fallback when direct paths aren’t available, based on the ICE framework. Trust between agents is bilateral and cryptographic.
pilotctl daemon start
pilotctl handshake agent-b
pilotctl send-message agent-b --data 'process this' --wait
Round-trip in my tests: 50 to 200ms via relay (direct peer-to-peer when paths allow: 10 to 30ms). Higher than gRPC or WebSockets. The tradeoff is that both agents can be anywhere — behind NAT, on different cloud providers, on a developer laptop — and the address stays the same regardless. I’ve had the same agent running for 60 days across multiple network changes without manual reconfiguration, which I wrote about here.
The network also has 435 specialist data agents on Network 9 covering finance, weather, academic data, government records, and more. For agents that need live external data, the difference is substantial — benchmarks show about 12 seconds end-to-end versus 51 seconds scraping equivalent data through conventional APIs. No separate API keys per service.
The protocol is filed as an IETF Internet-Draft, so the spec is stable and public. There’s more on how it fits alongside MCP and A2A here.
Verdict: Higher latency than gRPC but no infrastructure to run, works anywhere including behind NAT, and the data network is a real accelerator.
The comparison
| Method |
Avg latency |
NAT traversal |
Infrastructure needed |
Setup complexity |
| HTTP polling |
poll interval + 50ms |
No |
None |
Low |
| Webhooks |
50-200ms |
No |
Public endpoint |
Low |
| WebSockets |
10-40ms |
No |
Server |
Medium |
| MQTT |
5-100ms |
No |
Broker |
Medium |
| gRPC |
5-15ms |
No |
Server |
High |
| Pilot Protocol |
10-200ms |
Yes |
None |
Low |
The NAT column is the one that catches people. Every option except Pilot Protocol requires at least one agent to have a publicly reachable address. For cloud-only deployments that’s manageable. For agents running on developer machines, edge nodes, or across providers, you’re either running infrastructure or you’re writing NAT traversal code yourself — which is not a small amount of work. I wrote about what that actually takes here.
What I actually use
For agents that both live in a controlled cloud environment and have stable IPs: gRPC if I need performance, WebSockets if I need bidirectional streaming with less schema overhead.
For anything that needs to work across environments, or where agents might be on laptops or edge devices, or where I don’t want to run a broker: Pilot Protocol. The latency is acceptable and the operational simplicity is real — no infrastructure to maintain, no NAT configuration, the same code works everywhere.
If you’re still figuring out the infrastructure question before picking a protocol, the 5-point checklist I use before any agent deployment is here.
