Agentic AI
Google Releases AX: A Kubernetes-Inspired Orchestrator for Autonomous AI Agents
Google has open-sourced AX, a declarative orchestrator designed to efficiently manage and scale autonomous AI agents by suspending idle workloads and resuming them in sub-second intervals.
InfoQ
Google has unveiled AX, an open-source orchestration platform under the Apache 2.0 license that addresses a fundamental mismatch between how traditional infrastructure runs workloads and how autonomous AI agents operate. The system treats agents as stateful, long-running entities rather than fitting them into conventional microservice or batch-processing models.
Why Agents Need Different Infrastructure#
Autonomous AI agents operate differently than typical software. They spend much of their execution time idle—waiting for model inference, external API responses, or human feedback—interspersed with bursts of intensive computation during reasoning and tool execution. Traditional Kubernetes orchestration handles this poorly: keeping containers active during idle periods wastes compute resources, while spinning down and restarting agents incurs cold-start penalties that slow interactive workflows. AX solves this by suspending agent execution state when idle and resuming it in under a second, without the overhead of container restarts.
How AX Manages Agent Workloads#
Built on top of Agent Substrate (Google's execution runtime), AX packages agents as isolated actor sandboxes with strict resource boundaries. When an agent enters an idle state, the platform checkpoints its execution context and suspends it. Multiple agents can time-share the same underlying compute hosts, multiplying resource efficiency. The system exposes four declarative, Kubernetes-style primitives:
- Task: Defines execution lifecycle and resource constraints
- Workspace: Handles pre-execution setup—mounting repositories, configuring Model Context Protocol servers, and bootstrapping dependencies
- Gateway: Manages network security and credential injection for outbound requests
- Model: Provides a unified control point for LLM provider settings and secrets
Operators interact with AX through a command-line tool (written in Go) offering commands like ax apply (register workloads), ax watch (stream task status), ax ssh (debug sandboxes), and ax suspend/resume (manual control).
