Multi-Agent Banking Flows on Kubernetes: LangGraph, a Registry and AgentGateway
Build FastAPI specialist agents, a LangGraph orchestrator with an LLM planner, deploy them to a cluster with a registry, then add routing and rate limits with AgentGateway.
This article is published in English.
MCP covers the agent-to-tool case: a fixed function performs one narrow job and returns its output, with no autonomy or judgment on the tool side. Agent-to-agent communication is a different thing. The party on the other end can reason for itself, possibly with its own LLM, its own tools and its own decisions, and one agent hands it a sub-task and expects a reasoned answer back, not a lookup.
A bank account-opening process shows the shape of the problem. It involves several independent concerns: identity verification, customer history, background and criminal checks, financial capability and public records. Each could plausibly belong to a different team, be built separately and eventually be reasoned about separately. That is the kind of problem agent-to-agent architecture exists for. The build below grows from one agent into a team of specialists, a LangGraph orchestrator with an LLM planner, a Kubernetes deployment, a registry and finally a gateway in front.
Five specialists, one pattern
The account-opening team has five specialist agents. Each is a self-contained FastAPI service on its own port, and the identity verification agent sets the template: plain FastAPI with one endpoint.
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI(title="Identity Verification Agent")
class VerifyRequest(BaseModel):
customer_id: str
document_type: str
document_number: str
class VerifyResponse(BaseModel):
customer_id: str
status: str
confidence: float
notes: str
@app.get("/health")
async def health():
return {"agent": "identity-verification", "status": "ok"}
@app.post("/invoke", response_model=VerifyResponse)
async def invoke(req: VerifyRequest):
return VerifyResponse(
customer_id=req.customer_id,
status="VERIFIED",
confidence=0.97,
notes=f"{req.document_type} {req.document_number} matched on file.",
)
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=9001)
Every agent in the system follows this same pattern: a Pydantic request model, a Pydantic response model, one /invoke endpoint and one /health endpoint. A quick check:
curl -s http://localhost:9001/health
curl -s -X POST http://localhost:9001/invoke \
-H "Content-Type: application/json" \
-d '{"customer_id":"CUST-5567","document_type":"passport","document_number":"X1234567"}'
The remaining services are built exactly the same way, and eight services exist by the end.
First orchestrator: parallel calls
The first orchestrator version called all eight agents concurrently with asyncio.gather and synthesized a decision from the combined results:
async def call_agent(client, name, payload):
try:
resp = await client.post(f"{AGENTS[name]}/invoke", json=payload, timeout=10.0)
resp.raise_for_status()
return resp.json()
except httpx.HTTPError as e:
return {"status": "ERROR", "error": str(e)}
@app.post("/open-account")
async def open_account(app_data: AccountApplication):
async with httpx.AsyncClient() as client:
identity, history, background, financial, public_records = await asyncio.gather(
call_agent(client, "identity", {...}),
call_agent(client, "customer_history", {...}),
call_agent(client, "background_check", {...}),
call_agent(client, "financial_capability", {...}),
call_agent(client, "public_records", {...}),
............
............
)
# decision synthesized from all five results
The limitation pushed the design forward. Every agent got called every time, even when an early check had already failed. Some checks cost actual money, since background and public-records lookups usually come from paid outside providers, so the repeated calls are plain waste. The code also didn't separate the internal reason for a decision from the message shown to the customer. Error handling came down to one generic try/except wrapped around all the calls instead of something auditable per step.
The fix is a real LangGraph orchestrator, where each step can inspect what came before and route accordingly. The helper that calls an agent never raises. Every failure becomes status='ERROR', so the graph routes failures uniformly instead of crashing:
async def call_agent(agent_name: str, payload: dict) -> dict:
"""Never raises -- every failure mode becomes status='ERROR' so the
graph can route uniformly instead of crashing."""
url = f"{AGENTS[agent_name]}/invoke"
try:
async with httpx.AsyncClient(timeout=TIMEOUT_SECONDS) as client:
resp = await client.post(url, json=payload)
resp.raise_for_status()
return resp.json()
except httpx.TimeoutException:
return {"status": "ERROR", "error": f"{agent_name} agent timed out after {TIMEOUT_SECONDS}s"}
except httpx.ConnectError:
return {"status": "ERROR", "error": f"{agent_name} agent is unreachable"}
except httpx.HTTPStatusError as e:
return {"status": "ERROR", "error": f"{agent_name} agent returned {e.response.status_code}"}
except Exception as e:
return {"status": "ERROR", "error": f"{agent_name} agent call failed: {e}"}
def route_after_identity(state) -> str:
result = state.get("identity_result") or {}
if result.get("status") == "ERROR":
state.setdefault("reasons", []).append(f"Identity check failed: {result.get('error')}")
return "decline"
if result.get("status") != "VERIFIED":
state.setdefault("reasons", []).append("Identity could not be verified.")
return "decline"
return "continue"
# ... one routing function per check, same shape ...
builder = StateGraph(ApplicationState)
builder.add_node("identity_check", identity_check)
builder.add_node("customer_history_check", customer_history_check)
builder.add_node("background_check", background_check)
builder.add_node("financial_capability_check", financial_capability_check)
builder.add_node("public_records_check", public_records_check)
builder.add_node("decline", decline_node)
builder.add_node("approve", approve_node)
builder.add_edge(START, "identity_check")
builder.add_conditional_edges("identity_check", route_after_identity,
{"continue": "customer_history_check", "decline": "decline"})
# ... same pattern chained through all five checks ...
builder.add_edge("decline", END)
builder.add_edge("approve", END)
graph = builder.compile()
This is a straightforward LangGraph solution with error handling woven into each step. It still lacks the complexity real flows demand.
One LLM planner for many products
Account opening is only one banking product. A bank also handles credit cards, debit cards and gift cards, and each calls for a different set of checks. Hardcoding an if/else per product type doesn't scale, so an LLM planner (a GPT model through Azure OpenAI) takes a plain-English enquiry together with a catalog of what each agent offers, then picks the relevant agents and their order. That is reasoning, not branching logic.
Three product-specific agents joined the original set, which gives three flows: opening a bank account, linking a debit card and purchasing a gift card.
# A simple starter for all agents
uv run agents/background_check_agent.py &
uv run agents/card_linking_agent.py &
uv run agents/identity_agent.py &
uv run agents/customer_history_agent.py &
uv run agents/public_records_agent.py &
uv run agents/financial_capability_agent.py &
uv run agents/gift_card_compliance_agent.py &
uv run orchestrator_langgraph_llm.py &
chmod a+x ./start_agent.sh
# Run it
./start_agents.sh
(account-opening-agents) krishnansriram@Krishnans-MacBook-Pro account-opening-agents % INFO: Started server process [22183]
INFO: Started server process [22184]
INFO: Started server process [22185]
INFO: Started server process [22182]
INFO: Started server process [22186]
INFO: Started server process [22187]
INFO: Started server process [22188]
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:9007 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9001 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9005 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9008 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9003 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9002 (Press CTRL+C to quit)
INFO: Uvicorn running on http://0.0.0.0:9004 (Press CTRL+C to quit)
INFO: Started server process [22190]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:9000 (Press CTRL+C to quit)
The executor became one node that loops through whatever plan the LLM produced and no longer follows a fixed chain of graph edges. It was checked directly, bypassing the real LLM call since that needs live Azure credentials, by feeding it hand-built plans:
- A gift-card plan with one step,
gift_card_compliance, was approved under the anti-money-laundering threshold and declined over it, with nothing else called. - A three-step credit-card plan of
identity,financial_capabilityandcredit_limitran all three in the right order with the right results.
A small local script starts every agent at once, saved as start_agent.sh:
curl -s -X POST http://localhost:9000/process-enquiry \
-H "Content-Type: application/json" \
-d '{
"enquiry_text": "I would like to open a new checking account",
"customer_id": "CUST-9003",
"full_name": "Jordan Alex Smith",
"date_of_birth": "1990-04-12",
"document_type": "passport",
"document_number": "X1234567",
"declared_annual_income": 95000,
"address": "123 Main St, Columbus, OH"
}' | python3 -m json.tool
Handling connection for 9000
{
"customer_id": "CUST-9003",
"product_type": "account_opening",
"planned_steps": [
"identity",
"financial_capability",
"background_check",
"public_records"
],
"planner_reasoning": "This is a new deposit account opening request. Identity should be verified first for the new customer, followed by financial capability checks needed for account opening. Background screening and public records checks are also required for opening a new deposit account.",
"decision": "APPROVED",
"customer_message": "Your account opening request has been approved.",
"reasons": [
"All required checks passed."
],
"step_results": {
"identity": {
"customer_id": "CUST-9003",
"status": "VERIFIED",
"confidence": 0.97,
"notes": "passport X1234567 matched on file."
},
"financial_capability": {
"customer_id": "CUST-9003",
"status": "PASS",
"income_verified": true,
"estimated_credit_score": 742,
"affordability_status": "ADEQUATE"
},
"background_check": {
"customer_id": "CUST-9003",
"status": "PASS",
"criminal_record_found": false,
"sanctions_hit": false,
"watchlist_hit": false,
"notes": "No adverse findings for Jordan Alex Smith."
},
"public_records": {
"customer_id": "CUST-9003",
"status": "PASS",
"address_verified": true,
"bankruptcy_history": false,
"litigation_history": false
}
}
}
With the agents running, send a test enquiry:
curl -s -X POST http://localhost:9000/process-enquiry \
-H "Content-Type: application/json" \
-d '{
"enquiry_text": "I would like to link a debit card to my checking account",
"customer_id": "CUST-9002",
"full_name": "Priya Nair",
"date_of_birth": "1994-03-08",
"document_type": "passport",
"document_number": "X5566778",
"declared_annual_income": 65000,
"address": "12 Maple Rd, Dublin, OH",
"linked_account_number": "ACC-99887766"
}' | python3 -m json.tool
Handling connection for 9000
{
"customer_id": "CUST-9002",
"product_type": "debit_card",
"planned_steps": [
"customer_history",
"card_linking"
],
"planner_reasoning": "This is a debit card request. First check customer history to confirm the customer relationship and whether identity/KYC is already established, then perform card linking to verify the checking account and link the debit card.",
"decision": "APPROVED",
"customer_message": "Your debit card request has been approved.",
"reasons": [
"All required checks passed."
],
"step_results": {
"customer_history": {
"customer_id": "CUST-9002",
"status": "PASS",
"existing_customer": true,
"relationship_years": 3.5,
"prior_accounts": 1,
"kyc_status": "CURRENT"
},
"card_linking": {
"customer_id": "CUST-9002",
"status": "PASS",
"account_verified": true,
"notes": "Account ACC-99887766 verified and eligible for debit card linking."
}
}
}
A second request exercises another flow, and a companion stop_agent.sh shuts everything down:
# This all that we need to kill our agents - stop_agent.sh
pkill -f "uv run agents/"
pkill -f "orchestrator_langgraph_llm.py"
chmod a+x ./stop_agent.sh
# Execute stop agent
./stop_agents.sh
INFO: Shutting down
INFO: Shutting down
INFO: Shutting down
INFO: Shutting down
INFO: Shutting down
INFO: Shutting down
INFO: Shutting down
(account-opening-agents) krishnansriram@Krishnans-MacBook-Pro account-opening-agents % INFO: Shutting down
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Finished server process [22185]
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Waiting for application shutdown.
INFO: Finished server process [22186]
INFO: Application shutdown complete.
INFO: Finished server process [22190]
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Finished server process [22182]
INFO: Waiting for application shutdown.
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Application shutdown complete.
INFO: Finished server process [22184]
INFO: Finished server process [22187]
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Finished server process [22188]
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
INFO: Finished server process [22183]
The agents work locally, but there is more to do. Next they have to run behind AgentGateway, with a registry service alongside.
Deploying in a cluster
Start with a namespace for all the agents, which makes later discovery easier. It is called banking-agents:
kubectl get namespace banking-agents 2>/dev/null || kubectl create namespace banking-agents
One Dockerfile for every agent
There are two Dockerfiles: one for the application agents, and one for the registry and the orchestrator. Because every agent is a plain FastAPI app with the same shape, a single file, Dockerfile.agent, serves them all:
FROM python:3.13-slim
COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
WORKDIR /app
COPY pyproject.toml uv.lock ./
RUN uv sync --frozen --no-dev --no-install-project
COPY agents ./agents
ARG APP_MODULE
ENV APP_MODULE=${APP_MODULE}
CMD uv run python ${APP_MODULE}
Deploying agents one at a time
Here is the full flow for the first agent, identity, followed by the same three commands for each of the others:
docker build -t identity-agent:v1 \
--build-arg APP_MODULE=agents/identity_agent.py \
-f Dockerfile.agent .
kind load docker-image identity-agent:v1 --name agentgateway
kubectl apply -f- <<EOF
apiVersion: apps/v1
kind: Deployment
metadata:
name: identity-agent
namespace: banking-agents
spec:
replicas: 1
selector:
matchLabels: {app: identity-agent}
template:
metadata:
labels: {app: identity-agent}
spec:
containers:
- name: identity-agent
image: identity-agent:v1
imagePullPolicy: IfNotPresent
ports: [{containerPort: 9001}]
---
apiVersion: v1
kind: Service
metadata:
name: identity-agent
namespace: banking-agents
spec:
selector: {app: identity-agent}
ports: [{port: 80, targetPort: 9001}]
type: ClusterIP
EOF
kubectl rollout status deploy/identity-agent -n banking-agents --timeout=60s
The pattern repeats for all eight agents, which listen on ports 9001 through 9008 and cover customer history, background check, financial capability, public records, credit limit, card linking and gift-card compliance. Only the name, image and port change each time. Then confirm the pods:
kubectl get pods -n banking-agents
NAME READY STATUS RESTARTS AGE
background-check-agent-566d595b77-h69zg 1/1 Running 0 19h
card-linking-agent-786f4899-96s6c 1/1 Running 0 19h
credit-limit-agent-7fb6db7b76-b8zzf 1/1 Running 0 19h
customer-history-agent-b765d95fd-2pzh5 1/1 Running 0 19h
financial-capability-agent-5c879c8879-98sgp 1/1 Running 0 19h
gift-card-compliance-agent-75c55c74d4-qzpx4 1/1 Running 0 19h
identity-agent-668c66fff8-stgxt 1/1 Running 0 19h
public-records-agent-66dfd78c8b-qrcvh 1/1 Running 0 19h
All eight should read 1/1 Running.
Deploying the registry
The registry service references every other service in the system. Until it is up, the agents have no map showing how they connect. Later, access to it should be locked down so only application agents can reach it, which keeps agents secure and gives them a clear way to talk to each other.
docker build -t agent-registry:v1 --build-arg APP_MODULE=agent_registry_service.py -f Dockerfile .
kind load docker-image agent-registry:v1 --name agentgateway
kubectl apply -f- <<EOF
apiVersion: apps/v1
kind: Deployment
metadata:
name: agent-registry
namespace: banking-agents
spec:
replicas: 1
selector:
matchLabels: {app: agent-registry}
template:
metadata:
labels: {app: agent-registry}
spec:
containers:
- name: agent-registry
image: agent-registry:v1
imagePullPolicy: IfNotPresent
ports: [{containerPort: 9100}]
---
apiVersion: v1
kind: Service
metadata:
name: agent-registry
namespace: banking-agents
spec:
selector: {app: agent-registry}
ports: [{port: 80, targetPort: 9100}]
type: ClusterIP
EOF
kubectl rollout status deploy/agent-registry -n banking-agents --timeout=60s
kubectl port-forward -n banking-agents svc/agent-registry 9100:80 &
Verify that it can reach the other eight agents by DNS name. The response is a map of all application services, serving as a registry that holds locations and also descriptions, ideally paving the way for skills:
curl -s http://localhost:9100/agents | python3 -m json.tool
#Execution results
Handling connection for 9100
{
"agents": [
{
"name": "identity",
"url": "http://identity-agent.banking-agents.svc.cluster.local:80",
"description": "Verifies a customer's identity documents (passport, license, etc). Needed any time a NEW customer's identity hasn't already been established.",
"input_schema": {
"customer_id": "customer_id",
"document_type": "document_type",
"document_number": "document_number"
}
},
{
"name": "customer_history",
"url": "http://customer-history-agent.banking-agents.svc.cluster.local:80",
"description": "Looks up existing relationship, prior accounts, and KYC status for a customer. Useful to check whether identity verification can be skipped for an existing customer.",
"input_schema": {
"customer_id": "customer_id"
}
},
{
"name": "background_check",
"url": "http://background-check-agent.banking-agents.svc.cluster.local:80",
"description": "Criminal record, sanctions, and watchlist screening. Required for opening a new deposit account; usually not required for issuing a card to an already-verified customer.",
"input_schema": {
"customer_id": "customer_id",
"full_name": "full_name",
"date_of_birth": "date_of_birth"
}
},
{
"name": "financial_capability",
"url": "http://financial-capability-agent.banking-agents.svc.cluster.local:80",
"description": "Verifies income and estimates a credit score. Required for account opening and for credit card applications; not required for debit or gift cards.",
"input_schema": {
"customer_id": "customer_id",
"declared_annual_income": "declared_annual_income"
}
},
{
"name": "public_records",
"url": "http://public-records-agent.banking-agents.svc.cluster.local:80",
"description": "Checks address verification, litigation, and bankruptcy history. Required for opening a new deposit account.",
"input_schema": {
"customer_id": "customer_id",
"address": "address"
}
},
{
"name": "credit_limit",
"url": "http://credit-limit-agent.banking-agents.svc.cluster.local:80",
"description": "Determines an approved credit limit based on income and the requested limit. Required ONLY for credit card applications.",
"input_schema": {
"customer_id": "customer_id",
"declared_annual_income": "declared_annual_income",
"requested_credit_limit": "requested_credit_limit"
}
},
{
"name": "card_linking",
"url": "http://card-linking-agent.banking-agents.svc.cluster.local:80",
"description": "Verifies a bank account to link a debit card to. Required ONLY for debit card applications.",
"input_schema": {
"customer_id": "customer_id",
"linked_account_number": "linked_account_number"
}
},
{
"name": "gift_card_compliance",
"url": "http://gift-card-compliance-agent.banking-agents.svc.cluster.local:80",
"description": "Checks a gift card purchase amount against AML limits. Required ONLY for gift card purchases.",
"input_schema": {
"customer_id": "customer_id",
"purchase_amount": "purchase_amount"
}
}
]
}
Deploying the orchestrator
Azure OpenAI credentials go in as a Secret and are never baked into the image:
kubectl create secret generic azure-openai-secret \
-n banking-agents \
--from-literal=AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com" \
--from-literal=AZURE_OPENAI_DEPLOYMENT="gpt-5-1-chat" \
--from-literal=AZURE_OPENAI_API_KEY="your-key" \
--from-literal=AZURE_OPENAI_API_VERSION="2025-04-14"
Then build and deploy the image, following the same pattern as every other service:
docker build -t orchestrator:v1 --build-arg APP_MODULE=orchestrator_dynamic.py -f Dockerfile .
kind load docker-image orchestrator:v1 --name agentgateway
kubectl apply -f- <<EOF
apiVersion: apps/v1
kind: Deployment
metadata:
name: orchestrator
namespace: banking-agents
spec:
replicas: 1
selector:
matchLabels: {app: orchestrator}
template:
metadata:
labels: {app: orchestrator}
spec:
containers:
- name: orchestrator
image: orchestrator:v1
imagePullPolicy: IfNotPresent
ports: [{containerPort: 9000}]
env:
- name: AGENT_REGISTRY_URL
value: "http://agent-registry.banking-agents.svc.cluster.local:80"
envFrom:
- secretRef:
name: azure-openai-secret
---
apiVersion: v1
kind: Service
metadata:
name: orchestrator
namespace: banking-agents
spec:
selector: {app: orchestrator}
ports: [{port: 80, targetPort: 9000}]
type: ClusterIP
EOF
kubectl rollout status deploy/orchestrator -n banking-agents --timeout=60s
kubectl port-forward -n banking-agents svc/orchestrator 9000:80 &
Time for a test:
curl -s -X POST http://localhost:9000/process-enquiry \
-H "Content-Type: application/json" \
-d '{
"enquiry_text": "I would like to purchase a $500 gift card",
"customer_id": "CUST-7001", "full_name": "Jordan Smith", "date_of_birth": "1990-04-12",
"document_type": "passport", "document_number": "X1234567",
"declared_annual_income": 95000, "address": "123 Main St, Columbus, OH",
"purchase_amount": 500
}' | python3 -m json.tool
Handling connection for 9000
{
"customer_id": "CUST-7001",
"product_type": "gift_card",
"planned_steps": [
"gift_card_compliance"
],
"planner_reasoning": "This is a gift card purchase, so the only relevant specialist is gift_card_compliance to check the amount against AML limits. No identity, credit, account, or background screening steps are needed for this request.",
"decision": "APPROVED",
"customer_message": "Your gift card request has been approved.",
"reasons": [
"All required checks passed."
],
"step_results": {
"gift_card_compliance": {
"customer_id": "CUST-7001",
"status": "PASS",
"within_aml_limit": true,
"notes": "Purchase amount 500.0 vs AML threshold 2000.0."
}
}
}
A working multi-agent system with an orchestrator now runs in the cluster, which is a long way from a single endpoint.
Putting the orchestrator behind the gateway
Why front the orchestrator with a gateway at all? There are five reasons:
- One stable entry point. A real URL such as
https://your-domain/orchestratorreplaces a tunnel only a Kubernetes admin can use. Any real client, from a mobile backend to another internal service or a partner integration, can reach it normally. - Auth before code. With the JWT and RBAC policy from the security work, an unauthenticated or unauthorized caller is rejected at the gateway and never touches the orchestrator's application logic.
- Throttling. A client that misbehaves, or a bug in some upstream caller, could flood the orchestrator with requests, and every one of them reaches the planning model and adds to the Foundry bill. The gateway can cap that.
- One place to see all traffic. Logging and observability for every call come from the gateway instead of instrumentation inside the orchestrator.
- A stable surface when things change. Swap orchestrator versions, add a replica or move to another Service, and callers notice nothing, since they only ever knew the gateway path.
The earlier banking MCP servers made this same point for tools. The argument carries up one level, since the front door of a whole multi-agent system deserves the governance the individual agents got.
A plain HTTP backend such as the orchestrator, with no MCP translation and no AI provider config, doesn't need an AgentgatewayBackend. An HTTPRoute can point straight at a Kubernetes Service. One wrinkle remains: the orchestrator's Service lives in banking-agents while the HTTPRoute lives in agentgateway-system, a different namespace. Gateway API demands an explicit ReferenceGrant in the target namespace before allowing a cross-namespace reference.
First, allow the cross-namespace reference:
kubectl apply -f- <<EOF
apiVersion: gateway.networking.k8s.io/v1beta1
kind: ReferenceGrant
metadata:
name: allow-agentgateway-to-orchestrator
namespace: banking-agents
spec:
from:
- group: gateway.networking.k8s.io
kind: HTTPRoute
namespace: agentgateway-system
to:
- group: ""
kind: Service
name: orchestrator
EOF
Second, route directly to the Service:
kubectl apply -f- <<EOF
apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
name: banking-orchestrator-route
namespace: agentgateway-system
spec:
parentRefs:
- name: agentgateway-proxy
rules:
- matches:
- path: {type: PathPrefix, value: /orchestrator}
backendRefs:
- name: orchestrator
namespace: banking-agents
port: 80
EOF
kubectl get httproute banking-orchestrator-route -n agentgateway-system
Then verify through the gateway:
curl -s -X POST http://localhost:8080/orchestrator/process-enquiry \
-H "Content-Type: application/json" \
-d '{
"enquiry_text": "I would like to purchase a $500 gift card",
"customer_id": "CUST-7001", "full_name": "Jordan Smith", "date_of_birth": "1990-04-12",
"document_type": "passport", "document_number": "X1234567",
"declared_annual_income": 95000, "address": "123 Main St, Columbus, OH",
"purchase_amount": 500
}' | python3 -m json.tool
# Execution Results
Handling connection for 8080
{
"customer_id": "CUST-7001",
"product_type": "gift_card",
"planned_steps": [
"gift_card_compliance"
],
"planner_reasoning": "This enquiry is for a gift card purchase. The only required specialist agent is gift_card_compliance to check the $500 amount against AML limits; no identity, background, or account-linking checks are needed for a straightforward gift card purchase.",
"decision": "APPROVED",
"customer_message": "Your gift card request has been approved.",
"reasons": [
"All required checks passed."
],
"step_results": {
"gift_card_compliance": {
"customer_id": "CUST-7001",
"status": "PASS",
"within_aml_limit": true,
"notes": "Purchase amount 500.0 vs AML threshold 2000.0."
}
}
}
# 2nd Execution
curl -s http://localhost:8080/orchestrator/health
Handling connection for 8080
# Execution results
{"agent":"dynamic-banking-orchestrator","status":"ok","registry_url":"http://agent-registry.banking-agents.svc.cluster.local:80"}
Built-in guardrails and rate limits
AgentGateway's built-in filters are called guardrails or prompt guards. They concern the content of LLM traffic and not generic REST payloads, so they apply to an /llm route, the Foundry connection, and not to the orchestrator's plain JSON.
The obvious candidate for a bank is PII masking on the LLM route. The orchestrator sends customer details such as names, dates of birth, document numbers, income and addresses to the planning model inside the prompt. As built, though, the agentic code doesn't use /llm for all model access. The gateway offers ready-made guardrails for sensitive data on that route, but this implementation calls the model through a REST-style path, so those protections never activate. The proper fix is to send LLM calls through /llm, a change still to come.
Rate limiting is a general traffic policy and applies directly to the orchestrator route:
kubectl apply -f- <<EOF
apiVersion: agentgateway.dev/v1alpha1
kind: AgentgatewayPolicy
metadata:
name: banking-orchestrator-ratelimit
namespace: agentgateway-system
spec:
targetRefs:
- group: gateway.networking.k8s.io
kind: HTTPRoute
name: banking-orchestrator-route
traffic:
rateLimit:
local:
- requests: 60
unit: Minutes
burst: 10
EOF
The burst: 10 setting is optional but useful. It allows short spikes above the steady 60 requests a minute, which helps a legitimate client that fires several requests close together, without changing the sustained limit. Check that the policy applied:
kubectl get agentgatewaypolicy banking-orchestrator-ratelimit -n agentgateway-system
NAME ACCEPTED ATTACHED AGE
banking-orchestrator-ratelimit True True 62s
Then try it:
for i in $(seq 1 75); do
curl -s -o /dev/null -w "%{http_code}\n" http://localhost:8080/orchestrator/health
done | sort | uniq -c
# Execution results
Handling connection for 8080
Handling connection for 8080
Handling connection for 8080
Handling connection for 8080
............................
............................
Handling connection for 8080
Handling connection for 8080
Handling connection for 8080
Handling connection for 8080
70 200
5 429
Roughly the first 70 requests, 60 steady plus the burst allowance, should return 200, with 429 responses after that in the same minute window. If everything returns 200, either the policy didn't attach properly, in which case kubectl describe agentgatewaypolicy banking-orchestrator-ratelimit -n agentgateway-system shows any targetRef resolution error, or 75 requests simply didn't exhaust the burst allowance fast enough. Raising the loop to 100 settles that.
That completes the agentic implementation. What remains is securing access and adding guardrails between the agents themselves.