Most AI assistants stop at generating text. Nova goes further — it can actually do things in the real world. Here's how our Action Layer makes that possible.
The Problem with Chat-Only AI
Large language models are incredible at understanding and generating text. But there's a fundamental gap between "I understand what you want" and "I'll do it for you." That gap is what we call the Action Layer.
Architecture Overview
The Action Layer sits between Nova's AI brain and the outside world. It has three core components:
**Intent Recognition**: Nova parses your request and identifies not just what you want, but the specific actions required. "Schedule lunch with Sarah" becomes: check calendar → find mutual availability → create event → send invitation.
**Safety Sandbox**: Every action goes through a safety check before execution. Destructive actions (deleting files, sending money) require explicit confirmation. Read-only actions (checking calendar, reading emails) execute immediately.
**Integration Adapters**: Each connected service has a dedicated adapter that translates generic actions into service-specific API calls. The Google Calendar adapter handles scheduling, the Hue adapter handles lighting, and so on.
How a Request Flows
1. You say: "Turn on the living room lights and play some jazz" 2. Intent Recognition identifies two actions: lights.on(room: living room) + music.play(genre: jazz) 3. Safety Sandbox classifies both as non-destructive → auto-approve 4. Hue adapter sends the API call to set living room lights to ON 5. Spotify adapter starts a jazz playlist 6. Nova confirms: "Living room lights are on and jazz is playing 🎶"
The entire flow takes under 2 seconds.
Error Handling
Real-world integrations fail. APIs go down, tokens expire, devices go offline. The Action Layer handles this gracefully:
- **Retry with backoff**: Transient failures get 3 retries with exponential backoff
- **Graceful degradation**: If Spotify is down, Nova tells you and offers alternatives
- **Rollback**: If a multi-step action partially fails, completed steps are rolled back where possible
What's Next
We're working on chaining actions into complex workflows, conditional logic ("if it's raining, skip the outdoor lights"), and user-defined custom actions. The Action Layer is designed to be extensible — adding new integrations is as simple as writing a new adapter.