Agent Substrate Introduces a Secure Runtime for Large-Scale Agent Sandboxes

Summary

Agent Substrate is an early-stage, secure-by-default execution runtime for running large numbers of stateful agent workloads in sandboxes. The project says it is designed to achieve up to 10 times the density of standard container runtimes, with resume operations below 500 milliseconds and more than 500 suspend/resume activations per second. It multiplexes many application-level actors onto a smaller pool of workers, exploiting the fact that agent-like workloads are often idle, while managing actor creation, suspension, resumption, worker assignment, and traffic routing. The runtime supports microVMs and gVisor, provides kernel and network isolation, and preserves volatile memory and filesystem state through full-state snapshots. Kubernetes supplies infrastructure provisioning and worker lifecycle management, while Substrate adds agent-specific scheduling and control. A demonstration multiplexes about 250 stateful actors across eight physical pods and reports more than 30x oversubscription in its sample setup. The project is framework-agnostic and describes integrations or compatibility with ADK agents, LangChain, Claude Code, CodeX, Antigravity, and MCP servers. It also provides examples including a distributed agent executor, kagent workloads, request parking, autoscaled worker pools, and stateful coding environments. The repository includes control-plane, node-supervisor, networking, snapshotting, CLI, benchmarking, and GCP provisioning components, with local kind and GKE development paths. Agent Substrate is not an SDK for building agents, and its workloads do not have to be AI agents, although agents are its primary target. The project is explicitly not an officially supported Google product, is in early development, is not ready for production use, and does not promise backward-compatible APIs.