I'm building an AI-native operating system for business, from first principles. It's called Simple (https://simple.dev).

With 20+ years of experience building an enterprise software platforms, I have seen firsthand that trying to bolt AI onto legacy systems is an architectural dead end. It’s like building a state-of-the-art 'smart penthouse' on top of a 100-year-old brick building. The foundation wasn't designed for the weight, the wiring can't handle the power demands, and you get a high-tech facade on a crumbling, inefficient core.

We decided to build the modern skyscraper from the ground up, designing the entire system around three core principles:

1. A Unified State Machine: We started with a single, transactional data model and a core set of APIs that can represent any business object or workflow. Everything from a customer record to an approval process is a primitive in this system.

2. Language as the Primary Interface: Natural language isn't just for Q&A; it's a first-class citizen for commands. A prompt like "Create an app to track sales leads with fields for status, deal size, and owner, then add a 3-step approval workflow for deals over $50k" directly executes against the core APIs, modifying the actual schema and logic in real-time. No consultants needed.

3. True Agentic Execution: Our AI agents are given credentials to this same core API layer. You can delegate multi-step, stateful tasks ("When a new lead is assigned, notify the rep on Slack, schedule a follow-up in my calendar for 3 days, and generate a draft outreach email using our template"). The agent executes this by making the same API calls a human developer would, but with the flexibility to handle variations.

For the nerds, here's the tech stack we're using to make this happen: The backend is built in Elixir; the BEAM VM's actor model and fault tolerance are perfect for managing thousands of concurrent agents and workflows. For performance-critical parts, we drop down to Rust via NIFs. Crucially, all custom logic — whether generated by an AI agent or a human — is compiled to WASM. This provides a secure, high-performance sandbox, giving us language flexibility and near-native speed for all automated tasks.

We're moving from a paradigm of "users hunting through menus" to "users delegating real work." It's an ambitious mission, and I'd love to hear what the HN community thinks of this philosophy and architectural approach.