AX: Designing for the Agent Experience

I remember the first time I read DX (for developer experience) and how "seen" I felt (tho we didn't say that in ~2011).

Now I wanna talk about my current approaches to design for AX the Agentic Experience.

Don’t take my word for it:

It’s 2026. Build. For. Agents. -Karpathy

Don't worry, a lot of UX and DX carries over. A fine rule of thumb would be - whats good for you, is good for the agent.

Notice Friction#

Naming#

I was building out a CLI and skill for my agents and on first use it totally borked calling it! (Im not sure exactly why I think it had to do with the similarity between its name and another skill I have AND in my new skill doc it references this other skill - build composable CLI's/skills ftw!).

So I just asked it, what did it think we should call it?

The names it came up with were cooler than the one I had too!

Problem solved.

Shape#

For a while after you build a CLI/skill, watch closely how it is used.

• Did the agent mistakenly mess up arguments? -> Go rewrite your CLI to match that.

• Did the agent try to use something that didn't exist? -> Go build it.

Generalizing and Longevity#

But will this work across both Claude and Codex? Will this work for Codex 5.x?

I think it will decently because...

These LLMs (of the last handful of months) have been RL'ed on bash and CLIs enough that "what makes a good CLI" is now built in as an expectation that they carry around.

Might as well build toward its taste now.