AX
An open-source declarative runtime system developed by Google for AI Agents, written in Go and released under the Apache 2.0 license. It provides capabilities including persistent execution, sub-second crash recovery, and trajectory branching. The system is positioned similarly to Kubernetes but is designed specifically for Agent workloads.
Timeline (last 90 days)
Google has launched AX (Agent Executor), a runtime system for AI Agents.
AX is open-sourced under the Apache 2.0 license, and version v0.3.0 has been released.
AX originates from Google DeepMind research and Google's production experience in planetary-scale isolation and scheduling, and was released by engineer Jana Dogan.
The AX repository is located at google/ax, written in pure Go, has over 3,200 stars, and only 7 open issues.
AX can recover a dead Agent from a kernel snapshot after a crash in less than one second.
AX is composed of four declarative primitives: Task, Workspace, Gateway, and Model.
AX's CLI commands mimic kubectl, including ax apply, ax get tasks, ax watch, and ax ssh.
AX provides a Trajectory Branching feature that derives independent Agent instances by copy-on-write forking of kernel snapshots to explore different paths.
Suspended Agents consume no compute and incur zero cost during the wait period, making Human-in-the-loop patterns such as overnight approval queues economically viable.
The AX project is named after the axolotl, because of its ability to regenerate lost limbs, symbolizing an Agent regenerated from a snapshot.
2 more timeline events
All Facts (10)
AX is compatible with LangChain, LangGraph, ADK, and the A2A protocol, and federates Google's own Antigravity, Deep Research, and Gemini API hosted Agents.
50%UnverifiedSuspended Agents consume no compute and incur zero cost during the wait period, making Human-in-the-loop patterns such as overnight approval queues economically viable.
50%UnverifiedThe AX project is named after the axolotl, because of its ability to regenerate lost limbs, symbolizing an Agent regenerated from a snapshot.
50%UnverifiedAX originates from Google DeepMind research and Google's production experience in planetary-scale isolation and scheduling, and was released by engineer Jana Dogan.
50%UnverifiedAX can recover a dead Agent from a kernel snapshot after a crash in less than one second.
50%UnverifiedAX is composed of four declarative primitives: Task, Workspace, Gateway, and Model.
50%UnverifiedAX's CLI commands mimic kubectl, including ax apply, ax get tasks, ax watch, and ax ssh.
50%UnverifiedAX provides a Trajectory Branching feature that derives independent Agent instances by copy-on-write forking of kernel snapshots to explore different paths.
50%UnverifiedAX is open-sourced under the Apache 2.0 license, and version v0.3.0 has been released.
50%UnverifiedGoogle has launched AX (Agent Executor), a runtime system for AI Agents.
50%