Your Agents Cannot Use What You Never Named Properly
A paper landed on arXiv this week about a problem that sounds like plumbing and turns out to be about judgment. A production agent harness has to look at a growing library of skills and pick the one that fits the task in front of it. At small scale there is no clever machinery involved. The planner simply reads the skill descriptions sitting in its prompt and chooses. The authors treat that in-context selection as a retrieval problem, and once you frame it that way, an uncomfortable conclusion follows. How a skill is represented decides whether it is ever found.
Most teams read that as a prompt engineering note and move on. I think it is the more interesting half of the story. Every company building on agents is quietly assembling a library of internal capabilities, and every one of those capabilities has to be described before anything can route to it. Description is not documentation here. It is the interface. A workshop full of good tools with the labels rubbed off is a workshop where nothing gets used, and the person who wrote the labels turns out to have been doing load bearing work all along.
A capability nobody described well is, as far as your systems are concerned, a capability you do not have.
This is where I see the repricing starting. The market has spent a decade paying for people who can build the thing, and paying rather less for the people who can say precisely what the thing does, where its edges are, and whether it is one capability or three. That second skill used to look like tidiness. Taxonomy work, naming conventions, API surface design, the discipline of writing an honest description of your own system. It was the first thing cut when a roadmap got tight. It is now the difference between an agent stack that finds its own tools and one that flails through a library it cannot read.
The people who are good at this are not a new category and they are not easy to spot from a CV. They are usually the senior engineer everyone routes design questions through, or the technical writer who kept asking what a service actually guaranteed. They were already rare. Now their output is what the routing layer runs on.
For hiring leaders, the practical read is simple. In your next platform or agent hire, put a real internal system in front of the candidate and ask them to describe its capabilities well enough that another system could choose between them. What comes back will tell you more than any take home build ever has.
Planning a senior AI or data hire and want the market view first? We share it either way.
Send us the brief →