Agent Substrate is a system built on top of Kubernetes that manages agent-like workloads with higher scale and efficiency than Kubernetes alone can provide, while reducing latency. Written in Go, it operates by taking the Kubernetes control plane out of the critical path, allowing it to run on any existing Kubernetes cluster without interfering with standard Kubernetes workloads. The system maps a larger set of actors, which are applications such as AI agents, onto a smaller set of ready Kubernetes Pods, leveraging the fact that agent-like applications tend to be idle most of the time to achieve heavy multiplexing of up to 30x or more oversubscription.
The core functionality of Agent Substrate includes managing an actor's lifecycle through create, destroy, suspend, and resume operations, assigning actors to workers in real time, and routing incoming traffic to them. The system provides instant session teleport capabilities with sub-second activation of suspended actors onto available workers, state persistence through full-state snapshots that preserve both volatile RAM and filesystem state across hibernation cycles, and agent swarm multiplexing that can juggle hundreds of stateful actor sessions across a small pool of physical pods. The repository demonstrates these capabilities through multiple demos, including a Counter Demo that multiplexes approximately 250 stateful actor sessions across 8 physical pods, a Sandbox Demo running Alpine Linux with arbitrary shell execution, a Claude Code Multiplex demo, and a Secret Agent demo highlighting zero-idle self-suspension and re-animation.
Agent Substrate is intentionally framework and agent harness agnostic, managing standard OCI containers at the kernel level via gVisor. It provides native support for Agent Development Kit compatible session identity and persistent working memory, serves as an ideal execution environment for long-running stateful LangChain agents and sandboxed tool-calling, supports high-density stateful coding environments like Claude Code and CodeX that preserve terminal and filesystem state, and can deploy secure sandboxed Model Context Protocol servers as Substrate Actors.
The project is currently in very early development and not ready for production use, with APIs almost guaranteed to change and no backward compatibility guarantees.
The system comprises several core components including ateapi as the control plane API server, atelet as a node-level DaemonSet supervising worker pods, atecontroller for reconciling custom resources, atenet for networking and routing, ateom-gvisor for sandboxed execution, podcertcontroller as a Kubernetes polyfill, and kubectl-ate as a CLI management tool. The project currently aims to support the latest stable Kubernetes release and the previous minor release. Community engagement occurs through the ate-dev Google Group, weekly Thursday meetings, and CNCF Slack channels for both users and developers.
