The Event Bus (agent-nerves)

The Autonomic ecosystem is fundamentally decentralized. The 9 core “organs” do not communicate with each other via direct HTTP calls or shared memory. Instead, they operate as entirely independent processes that communicate asynchronously via a central event bus.

This event bus is managed by the agent-nerves daemon, which acts as a bridge to the underlying NATS server infrastructure.

Why a Decoupled Pub/Sub Architecture?

Traditional AI agents are monolithic. When an agent wants to run a shell command and parse the output, the entire process happens synchronously within a single context loop. If the terminal command hangs, the agent hangs. If the context window fills up, the agent crashes.

By decoupling the architecture using a Pub/Sub event bus, we achieve:

1. Resilience: If the visual QA organ (agent-eyes) crashes due to an Out of Memory error, the execution organ (agent-muscle) is completely unaffected. agent-body will restart agent-eyes, which will simply reconnect to the bus and pick up exactly where it left off.
2. Scalability: You can run agent-muscle on a high-powered GPU rig in the cloud while running agent-brain on your local laptop. As long as both machines are connected to the same NATS cluster, they can communicate seamlessly.
3. Observability: Because every action is an event traversing the bus, you can easily tail the traffic (e.g., using agent-nerves stream) to see exactly what the entire system is doing in real-time.

How it Works

The communication flow relies heavily on the NATS JetStream persistence engine.

1. Publishing: When an organ completes a task, it publishes an event. For example, when agent-muscle finishes deploying a frontend application, it publishes a deploy.frontend.success event.
2. Subscribing: Other organs subscribe to relevant subjects. agent-eyes is constantly listening for deploy.frontend.success events.
3. Action: When agent-eyes receives the event, it knows the deployment is ready, so it automatically navigates to the URL and captures a baseline screenshot.
4. State Management: The agent-spine workflow engine observes all events on the bus. It uses these events to advance the state of the active Directed Acyclic Graph (DAG) and trigger the next sequential nodes.

JetStream Persistence

We rely on NATS JetStream to provide durable subscriptions.

If an organ is temporarily offline (perhaps it is being restarted by agent-body or a new version is being deployed), it will not miss any events. When it reconnects to the NATS cluster, JetStream will deliver the backlog of events that occurred while it was down, ensuring zero data loss and guaranteed at-least-once delivery.

Event Filters (JSON & WASM)

The agent-nerves CLI allows you to register sophisticated event filters.

These filters can be written in simple JSON schemas or compiled to WebAssembly (WASM) for high-performance parsing. Filters are pushed directly down to the NATS server, ensuring that organs only receive exactly the events they are programmed to handle, preventing network saturation and CPU waste.