Writing on agent architecture.

When discovery moves onto a model's conversation instead of a search result, a website stops being where people experience the work and becomes a source read on their behalf.

When whether something works depends on its design rather than its code, a green test suite can sit on top of a puzzle nobody can solve.

A tool built to fit one specific workflow gets worse, not better, as it's made more generic — which is why domain expertise resists the centralization that swallowed the open web.

Running several AI agents on one machine is a fleet-management problem, not a chatbot-configuration problem.

On-device inference is a second democratization of AI: the first put models in the cloud for everyone, this one removes the internet connection and subscription as prerequisites entirely.

Cost per outcome drops fastest when the way you use a tool and the tool itself improve at the same time.

A system built for exactly one user needs no multi-tenant permission model to exploit and carries no shared risk from someone else's breach.

When AI absorbs execution across every business function, the constraint that remains isn't capacity, it's judgment — deciding what's worth doing, since AI makes nearly everything feel achievable.

Explicit context you write down beats context an assistant infers from your email and calendar, because you can see, edit, and correct exactly what it knows.

Whether autonomous AI agents can actually generate revenue depends on separating what's shipping from what's still hype — most of what gets called an autonomous agent today is a demo, not a business.