AX is Google's declarative orchestration runtime for running agent tasks in isolated workspaces on Kubernetes, with CLI commands for applying, watching, pausing, resuming, and shelling into tasks.
AX is an open-source orchestration runtime for agentic workloads. It centers on three declared resource types: Task, Workspace, and Model, and uses Kubernetes-style manifests and a CLI to manage them. The repository positions AX as a high-throughput system intended to run very large numbers of autonomous agent tasks in a cluster.
The project targets workloads that do not fit neatly into either stateless microservices or ordinary batch jobs. According to the README, agent tasks accumulate state, need strong isolation, interact with model APIs and tool servers, and can loop in costly ways if they are not supervised. AX is meant to make those behaviors manageable through declarative controls and operational commands.
At a high level, users define work through YAML manifests and apply them to a Kubernetes-backed control plane. A Task declares an agentic job, a Workspace prewires resources such as Git repositories, MCP servers, and skill packages, and a Model selects which LLM the platform should use with credentials stored in Kubernetes secrets. AX then sandboxes the task through Agent Substrate, attaches the declared workspace, and provides commands to observe, pause, resume, and inspect the running task.
It is gaining attention because it sits at the intersection of two fast-moving areas: agentic software and Kubernetes-style platform orchestration. The README emphasizes declarative workflows, isolated execution, and lifecycle controls like suspend, resume, watch, and ssh, which are practical concerns for teams trying to run agents reliably at scale. The repository also notes that the core concepts and specifications are still being refined, which often draws attention from early adopters following an evolving platform.
The README does not name direct competitors, but it explicitly compares the experience to Kubernetes and builds on Agent Substrate for sandboxed execution. Based on that, the closest alternatives are Kubernetes-native orchestration patterns for workloads, or other agent runtime and sandboxing stacks that manage isolated execution, workspace setup, and model configuration. Within the repository itself, Agent Substrate is the only concrete external dependency called out.
AI-explained · grounded in each repo's README