substrate is a zero-trust sandboxed runtime for Kubernetes that suspends idle agents and resumes them in under 500ms. It targets 10x higher density than standard containers when running very large numbers of mostly-idle, untrusted agents.
Project overview
It addresses a specific operational problem — running very large numbers of mostly-idle, untrusted agents securely and densely — by suspending idle agents and resuming them in under 500ms with default-deny networking and per-actor microVM sandboxes.
Project type
AI Agent · Infrastructure
Use cases
Automation
Deployment
Refer to project documentation
License
Apache-2.0
Best for
Operators who need to run very large numbers of mostly-idle, untrusted agents securely and densely on Kubernetes.
Users who can provide agent workloads as standard OCI container images and accept microVM sandboxing instead of plain Linux containers.
Key capabilities
Every actor runs in a dedicated gVisor or Cloud Hypervisor microVM sandbox with segregated compute, network, and storage. Plain Linux containers are not supported.
Default-deny networking with a built-in egress gateway that injects credentials into approved outbound requests outside the sandbox, so compromised agents have nothing to exfiltrate.
Limitations and risks
Plain Linux containers are not supported; every actor must run in a gVisor or Cloud Hypervisor microVM sandbox.
The quickstart requires Go, kubectl, and docker installed and multiple shell steps, so some command-line familiarity is needed.
Getting started
Install Go, kubectl, and docker; run hack/create-kind-cluster.sh; run go run ./cmd/ate-setup deploy ate-system; run go run ./cmd/ate-setup deploy demo counter; then create an actor and curl it via kubectl ate proxy.
Evidence and sources
README: a security-hardened, zero-trust runtime for sandboxed agents, engineered to safely run millions of agents and tools with 10x higher density than standard containers. Purpose-built…
README: Agent Substrate has reached 1.0. We are committed to compatibility and stability of the public API, CLI, and configuration formats.
README: It runs on standard Kubernetes: Pods, autoscaling, NetworkPolicies. It adds agent-specific scheduling, snapshotting, identity, and network mediation.
README: Every actor runs inside a dedicated gVisor (`runsc`) or Cloud Hypervisor microVM sandbox with segregated compute, network, and storage. Plain Linux containers are not supported.
README: The built-in egress gateway ([`atenet-egress`](docs/guides/egress-credential-injection.md)) injects credentials into approved outbound requests outside the sandbox, so a prompt-in…