# Agentic desire paths

A desire path is the unofficial shortcut pedestrians carve across a lawn.

Landscape architects used to treat these tracks as design failures, fighting against them with fences and plantings. But the better way to think about them is as data. Rather than fighting it, let the desire path emerge and then pave it.

Although it may look more unconventional, the design works better. It's built for how people actually want to use the space, not how you think they should.

If agents are going to use your product (and increasingly, they will) they're worth designing for, not just designing around. They have priors just like people do. Just like we instinctively reach for cmd+s to save, an agent reaches for yourcli status because it's seen git status a billion times, and it tries yourcli upgrade because that's what brew does.

When designing a CLI for humans, you want one way to do a thing, clearly documented. But if agents are your primary users, that may change. Every synonym for submit is a desire path. Every alias (push, send, publish, deploy) could be another route to the same destination, and the more of them you support, the less likely an agent gets stuck at a dead end.

Duplication that would feel sloppy in a human-facing tool becomes helpful redundancy in an agent-facing one.

I suspect we're doing the opposite of what we should to find these desire paths. Every CLI I see suggests stuffing its docs into your context window. But this just guides the agent to use the tool how we think it should.

If you're building a tool for agents, try stripping the context out and see what the agent reaches for on its own instead. What it attempts without instructions is demonstrated proof of how it thinks your tool should work and what patterns you should support.

The inclination when an agent misbehaves is to treat it as a prompting problem, refining the instructions until it does what you want. And sometimes that may be right.

But these models built their priors on billions of human <-> machine interactions.

What's more likely— that their intuition is wrong? Or that you are?