Most AI agent frameworks run on developer laptops. Gas City, an open-source orchestration SDK, treats Kubernetes as a first-class runtime — running each agent session as a Pod with native client-go integration, configurable resource limits, service accounts, and in-cluster service discovery.

This session presents Gas City's Kubernetes runtime provider. A declarative city.toml defines the agent fleet — providers (Claude Code, Codex, Gemini), resource profiles, and container images — while the controller reconciles desired state against running Pods. Each agent session runs in its own Pod with projected workspace volumes and injected Dolt (Beads database) service endpoints. When an agent's context window fills up, the Pod is recycled and a new one resumes the workflow molecule where the previous session left off.