Agent-Native Architectures

The number one principle of the author’s philosophy is that anything that a user can do through a UI an agent should be able to do through a sequence of tool calls. I agree with this and I think if every old and bespoke UI exposed such an atomic API the world would be a better place. Maybe the AI gold rush can force companies to make their apps more programmable.

The optimisation hierarchy in the article is also cool going from:

Agent composes atomic tools to achieve actions

Agent invokes domain specific skills

Agent flow is translated into code

Code is optimised in a lower level language

…

Assembly?

I like how it extends our normal notion of optimisation with new “agentic programs” which are nothing but model intelligence, simple but powerful tools and user intent.

Ultimately I couldn’t finish the article. “This isn't about a one-to-one mapping of UI buttons to tools—it's about achieving the same outcome”. Everytime I read contrast framing I take psychological damage and it really really hurts. Sometimes I feel like contrast framing is left in as a conspiracy by AI companies to easily make known people who are so lazy that they can’t even be bothered asking their agents to omit it. Like a watermark hidden in plain sight….

The article in fairness says that it was authored by Claude. And you can tell. When I first saw that the article was co-written by Claude I actually thought that it would be better thought out and structured. Surely, if somebody is brave enough to admit that they used AI for a blog, where the author’s voice is valued so highly, it must mean that they have cracked the code and produced something truly magnificent and they want to show it off! How wrong I was.