Agent Substrate is 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 for the era of autonomous agents, Substrate delivers fast resume operations in less than 500ms, with more than 500 activations per second. Agents get isolated compute, identity, networking, and storage. Sensitive credentials are kept out of reach from untrusted code. It supports multiple sandbox technologies, including microVMs and gVisor, while enabling a consistent lifecycle and management of all sandboxes.
At its core, Agent Substrate maps a larger set of “actors” (stateful workloads such as agents) onto a smaller pool of ready “workers”. Because agents spend most of their lives waiting on models, tools, or humans, Substrate snapshots idle actors out of memory and restores them when a request arrives.
Substrate is a low-opinion runtime, not an agent SDK. The workloads it runs do not have to be AI agents, though autonomous, untrusted, mostly-idle agents are the target for its isolation and scheduling model. It runs on standard Kubernetes: Pods, autoscaling, NetworkPolicies. It adds agent-specific scheduling, snapshotting, identity, and network mediation. This allows agentic, inference, training, and reinforcement learning workloads to share one infrastructure plane.
Substrate treats every actor as untrusted and mutually hostile, so tenants can share physical nodes without sharing a security boundary. Every control operates within the cold-start and resume budget. See the Security Guide for the full list of security principles and threat model.
Always sandboxed, VM-grade boundaries: 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.
Per-agent cryptographic identity: Every active actor receives a short-lived SPIFFE X.509 certificate by default, and the egress gateway can mint per-actor JWTs on demand when an egress policy calls for one. Both carry the actor identity and atespace (the Substrate namespace grouping), and are designed to federate with external identity systems.
Zero credentials in the sandbox: Actors never hold actor identity JWTs/certs, API keys, or cloud tokens. The built-in egress gateway (atenet-egress) injects credentials into approved outbound requests outside the sandbox, so a prompt-injected or compromised agent has nothing to exfiltrate.
Default-deny networking: Actors start with a private network stack, no inbound path except through the ingress router (atenet-router), no lateral actor-to-actor or control-plane access, and outbound traffic restricted by the egress gateway.
Pluggable policy at the edge: The egress gateway acts as an external Policy Enforcement Point (PEP) that is fully replaceable, allowing you to attach your own authorization, DLP, prompt-injection scanning, or cloud IAM policy engine without placing policy code inside the untrusted sandbox.
Watch the Agent Substrate cluster multiplex ~250 stateful actors across just 8 physical pods.
This demo highlights the core developer experience and "Agentic Infrastructure" capabilities of Substrate:
Actor Teleport: Suspend and resume actors onto any available worker in the pool, when needed.
Full-State Persistence: Working memory (RAM) and filesystem state are captured in snapshots and restored intact across hibernation cycles, without losing in-flight sessions or tool state.
10x–30x Oversubscription: Multiplex a large registry of stateful actors onto a small pool of warm worker pods, paying for compute only while an agent is actively thinking or executing a tool.
Agent Substrate is framework and agent harness agnostic. Because it runs standard OCI container images inside gVisor or microVM sandboxes and enforces security outside the workload, it hosts agents built on any stack without SDK changes:
Agent Development Kit (ADK): Preserves session state across invocations as actor state.
LangChain: Drop-in execution environment for LangChain agents and tool calls.
Claude Code, Codex, and Antigravity: High-density, stateful coding environments that keep system and filesystem state across sessions while isolating untrusted code execution.
Model Context Protocol (MCP): Deploy sandboxed MCP servers as Substrate Actors with per-actor identity and egress controls.
kagent: A CNCF Sandbox project and Kubernetes-native framework for building, deploying, and managing AI agents that uses Agent Substrate to run sandboxed, stateful agent workloads (see the announcement blog).
Agent Substrate has reached 1.0. We are committed to compatibility and stability of the public API, CLI, and configuration formats. We will continue to add new features, but we will not break existing APIs or configuration formats without a major version bump. See "Compatibility and schema evolution" in CONTRIBUTING.md for more details.
Please see CONTRIBUTING.md for guidelines on contributing to
the project. We welcome contributions of all kinds, but the project is VERY
young. Our immediate focus is on building out the core system and demos, so we
may not be able to review or merge contributions that don't align with those
goals in the near term.
Make sure you have Go, kubectl, and docker installed and configured on your dev machine. We will automatically manage other dependencies via Go, including kind.
A Linux host that will run more than about a thousand sandboxes also needs a larger kernel neighbor table (net.ipv4.neigh.default.gc_thresh1/2/3 = 4096/16384/32768); hack/create-kind-cluster.sh warns when the host is at the default of 1024.
Run the following steps:
# create cluster and local registry
hack/create-kind-cluster.sh
# make ate-setup target the kind cluster; set this in every shell that runs# ate-setup, including for the demo walkthroughsexport ATE_INSTALL_KIND=true
# install Substrate
go run ./cmd/ate-setup deploy ate-system --credential-provider='{"name":"k8s.io"}' \
--ingress-allowed-spiffe-ids=spiffe://cluster.local/ns/ate-system/sa/atenet-client
# install counter demo
go run ./cmd/ate-setup deploy demo counter
# install kubectl-ate
go install ./cmd/kubectl-ate
# create a counter actor in the demo's atespace (--template names the# actor template, resolved in the actor's atespace)
kubectl ate create actor my-counter-1-a ate-demo-counter --template counter
# serve a local proxy on port 8000 that sends requests to the actor through the# network router
kubectl ate proxy actor my-counter-1-a ate-demo-counter
[!IMPORTANT]
{"name":"k8s.io"} deploys the bundled credential provider with an empty,
default-deny namespace policy. Before an actor can have a credential
injected, grant its atespace the namespaces whose Secrets it may read; see
Enable it.
In a separate terminal, send an HTTP request to increment the counter:
You can deploy or cleanup specific Agent Substrate components using
the installation script. See go run ./cmd/ate-setup for all options.
# Re-deploy only ate-apiserver of the ATE system
go run ./cmd/ate-setup deploy apiserver
# Delete everything (core system and all demos)
go run ./cmd/ate-setup delete all
To deploy on Google Kubernetes Engine, follow the official
GKE Installation Guide.
It lists the prerequisites and gives the one-line installer command. The
installer's source and detailed options are in
substrate-gke.
We provide several sample applications demonstrating Agent Substrate's capabilities:
Counter Demo: A stateful Go HTTP server demonstrating state preservation across suspends/resumes, and on-demand actor resumption and routing via the Substrate router.
Sandbox Demo (Antigravity): A secure, sandboxed execution environment (running Alpine Linux) that allows arbitrary shell execution while preserving filesystem state across sessions.
Claude Code Multiplex: Demonstrates oversubscribing physical hardware by multiplexing multiple Claude Code agents onto a limited pool of workers.
Multi-Template: Two ActorTemplates running different binaries share one WorkerPool, even though the templates live in different atespaces.
Request Parking: An oversubscribed pool where the router holds inbound requests until a worker frees up, instead of returning 503.
