One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

agent-substrate/substrate

1,275+26 todayGo

Google's open-source system for cramming hundreds of AI agents onto just a handful of servers

Agent Substrate packs many mostly-idle AI agent processes onto a small pool of shared servers to save resources. Built on Kubernetes, it can suspend and resume an agent in under a second, letting it fake having far more workers than it actually has. In a demo, it ran 250 stateful agents on only 8 physical pods.

What it does

  1. It registers far more agents ('actors') than there are actual worker servers, exploiting the fact that agent-like apps are idle most of the time to multiplex them onto a small pool
  2. Agents run inside isolation technologies like microVMs or gVisor, letting the system suspend and resume them in under a second while preserving RAM and filesystem state via full snapshots
  3. It's not an SDK for writing agents but infrastructure for running already-built agents at scale, regardless of the framework (ADK, LangChain, Claude Code, MCP servers, etc.)
  4. It layers agent-specific scheduling on top of Kubernetes Pods and autoscaling, and its demo shows 250 stateful actors running across just 8 physical pods
  5. The project states it is in early development, not ready for production, and its APIs are expected to change

Why it matters

Running LLM-based agents in production usually means each one ties up server resources continuously, which drives costs up fast; this project tries to change that math by exploiting the fact that agents spend most of their time idle. Because it works at the container level rather than locking users into one agent framework, it's directly relevant to teams that need to run many different agent stacks at scale.

Terms in this repo

  • gVisor · A container sandbox technology from Google that isolates execution at the kernel level
  • microVM · A lightweight virtual machine that boots much faster and lighter than a full VM
  • actor · An individual agent/application instance managed by the system
  • worker · The physical resource unit (a Kubernetes Pod) that actually runs an actor
  • oversubscription · Registering more workloads than there is physical capacity, assigning resources only as needed

Repository description (English)

Agent Substrate: the core system

Open on GitHub

Trending repos

All repos →

Latest from METAL LAB