A control plane for agents that survive their own crashes

What it does

AX is a distributed agent runtime written in Go. It runs a single-writer controller that dispatches work to isolated local or remote actors—agents, skills, tools—and logs every step to a durable event log. If a client drops, a node reboots, or an actor crashes, the controller can resume execution from the last known state rather than starting over. It is compute-agnostic but targets Kubernetes via Agent Substrate for production deployments.

The interesting bit

The project treats resumption as a first-class primitive, not an afterthought. You can fork an existing conversation from any checkpoint into a new one, and the CLI supports catching up a disconnected client by replaying missed events from a sequence number. That design shifts the burden of reliability from the agent logic to the infrastructure layer.

Key highlights

• Single-writer architecture with a durable event log for automatic recovery
• Supports remote agents via gRPC, allowing tools and skills to run as isolated actors
• Built-in Gemini agent and bash tool with user-approval safety flow
• Web UI tracing via ax trace to visualize execution logs
• Aims for best experience on Kubernetes, though it runs anywhere

Caveats

• Explicitly in active early development with major breaking changes expected before any stable release
• External pull requests are temporarily paused while the core architecture stabilizes
• Documentation is incomplete; the README cuts off mid-sentence in the Custom Agents section

Verdict

Worth watching if you are building long-lived autonomous agents and need infrastructure-level crash recovery. Skip it if you want a polished, stable framework today.