Field notes · May 20, 2026

Designing for uncertainty

AI-native interfaces are probabilistic. Confidence, provenance and a graceful path to correction are now core UI primitives.

Deterministic software promises the same result every time. AI-native software does not, and pretending otherwise is the fastest way to lose a user's trust.

We treat three things as primitives: provenance (where did this come from), confidence (how sure is the system) and correction (how cheaply can I fix it). Every agent surface we design answers all three.

The output is not a chat box. It is a set of states — proposed, reviewed, accepted, reverted — designed as carefully as any checkout flow.

Why I joined Musebit AI

After a decade designing assistants inside large organisations, the interesting work moved to small, hyper-focused teams. Here's why.

Agentic workflows, in practice

A conversation on where autonomy actually pays off, and where a human checkpoint is worth every second it costs.