google/ax

An open source distributed agent runtime

What it solves

AX addresses the infrastructure challenges of running autonomous, long-running AI agents. Traditional runtimes are often designed for stateless services or batch jobs, making them inefficient for agentic workloads that are bursty—alternating between intense computation and long periods of idleness (e.g., waiting for human approval). AX provides a distributed runtime that manages state, ensures reliability, and allows agents to be suspended and resumed without losing progress.

How it works

AX acts as a distributed harness runtime that dynamically provisions isolated environments from images to execute agents and harnesses. It uses a single-writer architecture with a durable event log to maintain a consistent state and enable automatic recovery from failures. When deployed on compatible platforms like Agent Substrate on Kubernetes, it can suspend and resume compute-layer actors, allowing for high-density agentic workloads.

Who it’s for

It is designed for developers and researchers building sophisticated agentic applications who want to move away from monolithic agents toward distributed harnesses where tools, skills, and agents are deployed as isolated actors.

Highlights

  • Distributed Runtime: Executes harnesses, skills, and agents in isolation.
  • Native Resumption: Automatic recovery from interruptions and support for suspending/resuming stateful actors.
  • Durable Execution: Uses an event log to ensure state management and reliability.
  • Extensibility: Supports custom environment configurations, MCP tools, and custom harness implementations via HarnessService.
  • Kubernetes Optimized: Specifically designed to provide a high-density experience for agentic workloads on Kubernetes.

Related

  • Project
    agent-substrate/substrate

    Agent Substrate is a Kubernetes‑native runtime that multiplexes thousands of stateful AI agents onto a few sandboxed worker pods, offering sub‑500 ms resume times, full memory/filesystem snapshotting, and zero‑trust isolation via gVisor or micro‑VMs. It’s framework‑agnostic, provides a CLI (`kubectl‑ate`), and includes demos and integrations (Agent Executor, kagent). The project is early‑stage and not yet production‑ready.

  • Project
    agentspan-ai/agentspan

    A durable runtime for AI agents that prevents execution loss during process crashes and supports long-running, event-driven workflows with human-in-the-loop approvals.

  • Project
    sandbaseai/sandbase-harness

    A local-first runtime for AI agents that provides sandboxed code execution, persistent session management, and tool governance on the user's own infrastructure.

  • Project
    SWE-agent/SWE-ReX

    A runtime interface for AI agents to execute commands in sandboxed shell environments across various local and remote platforms.

  • Project
    greyhaven-ai/autocontext

    A recursive self-improvement harness that helps AI agents iteratively improve their performance on tasks by running evaluations and distilling lessons into playbooks and datasets.