Agent Substrate

Tracked

Kubernetes-based system for managing agent workloads at scale, multiplexing many stateful actors onto fewer pods with sub-second activation and persistent state.

Author Agent Substrate Open Sourced 2026-05-13 Last Commit Unknown

Overview

Agent Substrate is a Kubernetes-layer system purpose-built for running AI agent workloads at scale. It multiplexes a large set of stateful "actors" (agents) onto a smaller pool of ready "workers" (Kubernetes Pods), leveraging the fact that agents are idle most of the time. It achieves 30x+ oversubscription with sub-second actor activation and full state persistence across hibernation cycles.

Key Features

  • Actor-to-worker multiplexing with 30x+ oversubscription on Kubernetes
  • Sub-second suspend/resume with full RAM and filesystem state persistence
  • Framework-agnostic — works with ADK, LangChain, Claude Code, and any OCI container
  • Sandboxed execution via gVisor for secure isolation
  • MCP server deployment as durable Substrate Actors

Use Cases

  • Running thousands of concurrent AI agents on minimal Kubernetes infrastructure
  • High-density stateful coding environments (Claude Code, Codex) with session persistence
  • Deploying sandboxed MCP tool servers as durable actors
  • Cost-efficient agent infrastructure for production agentic applications

Technical Details

  • Built on Kubernetes, bypasses control plane for low-latency scheduling
  • Uses gVisor for kernel-level container isolation
  • Compatible with Agent Executor (AX) for distributed agent runtime coordination
  • Supports standard Kubernetes autoscaling alongside agent-specific scheduling