If the interface is evaporating, the obvious question is: build what instead?
My answer is the boring layer. The parts that don’t care whether a human or an agent is driving.
The four pieces I keep coming back to
- Memory. An agent with no memory is a very fast intern with amnesia. LoomBrain is my bet here, a second brain agents can actually read from.
- Sandboxes. Agents need somewhere to act without breaking the world. That’s sandcaster: full agent, secure sandbox, one call.
- Reasoning. Not just plausible next tokens, but a sensible path. I’ve been collecting research-backed strategies for smaller models and routing between them.
- Routing. The right capability, at the right moment, without a human wiring it up.
Why the boring layer
Because it lasts. Interfaces churn every few years; the need for memory, safe execution, and good judgment doesn’t. The unglamorous infrastructure is what everything else stands on, and it’s still wide open.
I’d rather own a plain pipe that everyone needs than a beautiful screen that’s about to be automated away.