AX is a declarative orchestrator for running autonomous agent workloads at high throughput in a Kubernetes cluster. You declare an agentic task with its workspaces and model specification in a manifest, and AX sandboxes the task, wires up its workspace, and schedules it. Each task runs as a sandboxed actor on Agent Substrate. AX is in heavy development, with major breaking changes expected before a stable release.
It treats agents as a distinct kind of workload: they accumulate state, need strict isolation, call model APIs and tool servers, and can run up costs in a loop. Three manifest kinds cover this:
| Kind | Purpose |
|---|---|
| Task | Runs untrusted agent code in an isolated sandbox with CPU and memory limits |
| Workspace | Pre-wires Git repos, MCP servers, and skill packages so every agent starts warm |
| Model | Configures which LLM the platform uses, with credentials from a Kubernetes secret |
The command-line client follows the shape of kubectl, with apply, get, describe, watch, and delete, plus agent-specific commands. Suspend checkpoints an idle task and resume picks it up where it left off, and an SSH command opens a shell inside a running sandbox when debugging is enabled. The CLI follows the active Kubernetes context and reaches the control plane over gRPC.
Features
- Declarative manifests: Task, Workspace, and Model kinds applied with one command
- Sandboxed tasks: each task runs as an isolated actor on Agent Substrate with resource limits
- Warm workspaces: Git repositories, MCP servers, and skill packages wired in before the agent starts
- Suspend and resume: checkpoint an idle agent and continue from the same state
- Sandbox shell: open a shell or run a one-off command inside a running task
- Live status: stream task phase and condition changes
- Context-aware CLI: follows the active Kubernetes context and tunnels to that cluster's control plane
