Run millions of idle autonomous agents at high density
Agent Substrate is a runtime designed for agents that spend most of their time in an idle state. It allows you to run millions of agents at a density ten times higher than standard containers. It operates on Kubernetes and provides each agent with isolated compute, network, and identity.
Removing idle time from memory
Autonomous agents spend the majority of their time waiting. Agent Substrate leverages this characteristic by storing idle agents (Actors) as snapshots in memory and restoring them to a Worker only when a request arrives. This approach is demonstrated by running approximately 250 stateful agents simultaneously on eight physical nodes. You pay for compute only when the agent is actually thinking or executing tools. This process occurs within a gVisor or Cloud Hypervisor microVM sandbox; standard Linux containers are not supported.
Security design for untrusted code
The system treats all agents as hostile to one another. Consequently, each agent runs only within a dedicated sandbox, which contains no sensitive credentials such as API keys or cloud tokens. Outbound traffic is filtered through an egress gateway, and credentials are injected only when necessary. Each agent receives a short-lived SPIFFE X.509 certificate to prove its identity, designed to integrate with external identity systems. The network is blocked by default, allowing access only through an ingress router.
Compatibility with existing agent frameworks
Agent Substrate is a runtime that does not depend on a specific agent SDK. Because it runs standard OCI container images, it can execute agents built with various stacks, including LangChain, ADK, Claude Code, Codex, and Antigravity, without modification. Model Context Protocol (MCP) servers can also be deployed as sandboxed Actors, applying individual identity and egress controls. This allows inference, learning, and reinforcement learning workloads to share the same infrastructure plane.
Things to check before adoption
Agent Substrate is currently a pre-1.0 version, and its API and behavior may change significantly. It supports the latest stable version of Kubernetes and the previous minor version. You may need to adjust the Linux kernel's neighbor table settings for large-scale sandbox execution. The project's current focus is on building the core system and demos, so contributions that do not align with this goal may not be merged. Before adopting it, you must verify whether beta APIs need to be enabled when creating a Kubernetes cluster and review the GCP resource provisioning process.
