Scaling Agents on Kubernetes with acpx and ACP: Isolation, Routing, and Cost per Agent

A 55-second walkthrough of how acpx and ACP package agent runtimes on Kubernetes: why the unit of work is a session rather than a request, where the gateway belongs, what has to live outside the pod, and why a prompt is not safe to retry blindly. What the video covers: - The ACP wire exchange: session/new, then session/prompt scoped by sessionId - Why a stdio JSON-RPC protocol needs a gateway inside each pod - Conversation, workspace and credential state that must survive pod eviction - The retry rule: only retry before the first session/update for a turn - Why scaling granularity is the session, not the request Honest caveat from the post: it has no access to the acpx runtime or a recording of the talk, so every load and cost figure is a placeholder with a stated method, not a benchmark. Run the probes yourself before you size anything. Full article: https://hackyjs.com/posts/scaling-agents-on-kubernetes-with-acpx-and-acp-isolation-routing-and-cost-per-agent

kubernetesai-agentsagent-infrastructureacpscaling

Read the post →