An agent architecture challenge
Build the agent fleet.
Defend every boundary.
Design a production support agent on Amazon EKS. Every choice changes latency, cost, isolation, or operational load.
Amazon EKS
Amazon EKS
Mission complete
Your agent architecture profile
Balanced platform
Constraint review
Would these boundaries hold?
Architecture record
Six decisions, one agent platform
Self-guided architecture lab
Stress the boundary. Inspect the consequence.
Contain model-generated code
Repair the isolation boundary without turning every request into a cold start.
Match the Kubernetes shape to the work
The same agent code can be a Service, Job, or session. Coordination changes independently.
lifecycle → primitive · topology → coordinator · duration → transport
Team shape does not dictate transport. A2A and MCP solve different directions.
A2A = agent to agent. MCP = agent to tool.
A Ready node is not a ready model
Watch scheduling, image pull, weight loading, readiness, and agent scaling complete in separate phases.
request resources → provision node → load model → pass readiness
Pods declare capabilities; Karpenter has not created matching nodes yet.
One model route changes. Which agents break?
Compare direct provider clients with one gateway endpoint as budget and availability change.
authenticate → budget → route → fallback
The primary route is healthy. Both designs serve traffic.
Choose where you want control and where you want offload
The options are peers, not maturity levels. Inspect the operating posture of each boundary.