AX

AX is Google's open-source agentic execution runtime, licensed under Apache 2.0 and hosted at github.com/google/ax. It provides a declarative, Kubernetes-style control plane purpose-built for autonomous agent workloads — a category the project describes as neither stateless microservices nor traditional batch jobs. The project was born from agentic runtime research at Google, drawing on work from Google DeepMind alongside large-scale isolation and scheduling experience, and runs on top of Agent Substrate, a compute runtime designed for massive actor density.

What It Is

AX is a high-throughput orchestrator that lets developers and researchers declare agent tasks as YAML manifests and run them at scale inside isolated sandboxes. It abstracts four core primitives — Task, Workspace, Gateway, and Model — into a unified control plane that handles sandboxing, workspace setup, network policy enforcement, and model configuration. If you have used Kubernetes, the README notes that ax will feel familiar: the CLI uses apply, get, describe, watch, and delete verbs, plus agent-specific commands like ax ssh, ax suspend, and ax resume.

Core Primitives

AX organizes everything around four declarative resource kinds:

• Task — runs untrusted agent code in a sandbox with CPU and memory limits; cheap to create, suspend, and discard.
• Workspace — pre-wires Git repositories, MCP servers, and skill packages so every agent starts warm; supports plain-English goal descriptions that an agent resolves on first boot.
• Gateway — defines and enforces network policies, locking traffic to an explicit allowlist of hosts and ports and injecting credentials into incoming requests.
• Model — centralizes LLM configuration, credentials (via Kubernetes secrets), and model parameters so rotating a key or pinning a new model version requires a single ax apply.

Architecture and Scale

AX runs on Agent Substrate, a compute runtime designed from the ground up for high actor density. The homepage states the platform is built to scale to billions of concurrent agent sessions per cluster. Key architectural properties include:

• Sub-second resumption — idle agents waiting on model responses, tool calls, or human approval are checkpointed and suspended, then brought back in under a second with no cold-start delay.
• Dense multiplexing — dozens of tasks share worker resources, converting idle waiting time into spare compute capacity.
• Generative workspaces — workspace goals described in plain English are handed to an agent on first boot to install toolchains and verify dependencies automatically.

Who It Is Built For

The project explicitly targets two audiences: application developers building production agentic systems and AI researchers. For researchers, AX is positioned as a platform for spinning up large numbers of reproducible sandboxes to collect trajectories, run reinforcement learning loops, and evaluate agents at scale. For developers, the design philosophy emphasizes ergonomics, rapid iteration, and minimal runtime surface area.

Update: v0.3.0

The latest release is v0.3.0, published on 2026-09-20, with the repository last pushed on 2026-09-25. The GitHub README includes an active warning that core concepts, protocols, and specifications are still being refined and that major breaking changes are likely before a stable release. A public roadmap covers planned milestones across core specs, actor architecture, agentic environments, and governance.