The Problem

Software today is designed for humans, not machines. That creates a structural mismatch:

• Automation and testing rely on brittle abstractions that break as soon as the UI changes
• Teams effectively maintain two parallel systems, one for humans (the UI) and one for machines (tests and automation)
• As AI accelerates software creation, verification and reliability are quickly becoming the bottleneck

At a fundamental level, computers don’t know how to use computers.

Why now

Multimodal models have reached the point where they can interpret UI snapshots reasonably well, but they still struggle with flow, memory, and interaction over time.

At the same time:

• AI is dramatically increasing the amount of software being built with the rise in AI tooling
• Verification and reliability are becoming critical bottlenecks
• Pure model wrappers are likely to commoditise, while data, feedback loops, and systems design become the real moat

The Opportunity

We are building agents that can understand and operate software interfaces in the same way humans do.

The goal is to make software interaction a learned capability, not something that has to be manually scripted, brittle, and re-built for every interface.

What we’re building

We are building a general capability for UI understanding and interaction, something closer to “learned usage” than traditional automation: