Agent Substrate — Summary

Agent Substrate is Google's open-source Kubernetes-based runtime for deploying AI agents at scale. Rather than running each agent as a persistent pod (which wastes resources on idle workloads), it multiplexes many "actors" onto a smaller pool of pre-started "worker" pods using sub-second suspend/resume via gVisor or micro-VM checkpoint/restore.

The system has six components: an ate-api-server (gRPC control plane backed by Redis/ValKey for high-frequency state), an atecontroller (Kubernetes controller reconciling WorkerPool and ActorTemplate CRDs), an atelet (node-level DaemonSet supervising snapshot lifecycle and OCI image preparation), an ateom (sandbox herder running inside each worker pod, driving runsc or Cloud Hypervisor), atenet (DNS + Envoy-based proxy for actor-aware routing and on-demand wakeup), and kubectl-ate (CLI).

Key architectural decisions: (1) dual-layer state model — CRDs for slow-changing config (WorkerPool, ActorTemplate) vs. Redis for fast-changing dynamic state (Actor, Worker records); (2) sandbox binaries not baked into worker images — fetched at runtime and pinned into snapshot manifests for reproducible restores; (3) self-describing snapshot manifests that include the exact sandbox binary versions; (4) concurrent download + OCI unpack during restore to minimize latency; (5) golden snapshots for template-level fast-start, plus DATA_ON_GOLDEN restore for micro-VMs where durable data is served atop a shared golden memory image.

North star metrics: 100ms P95 activation latency, 1 billion actors per cluster, 1,000 wakeups/second. The project is in early development (not production-ready, APIs will change).