Some of the most consequential technical work of the last forty years took the form of an agreement about how to describe something. Not a product, not a breakthrough, just a shared way of writing a date, naming a character, labelling a picture or asking one machine a question. It is hard to think of anything less likely to be celebrated, and hard to think of anything that has quietly carried more.
The appeal is easy to state. An agreement of that sort lets work done by people who will never meet fit together anyway. It lets a thing made this year be read in thirty years by something nobody has thought of yet. It lowers the cost of leaving, which turns out to be the most reliable guarantee of being treated well while you stay.
Having watched a number of these take hold and a number fail, the durable ones seem to share a short list of qualities:
- They are boring on purpose. A good agreement settles the dull question completely so that everybody can get on with the interesting one.
- They are small enough to implement in a week. If the full version takes a year to support, most people will support half of it, and half of an agreement is worse than none.
- They describe, rather than instruct. The lasting ones say what a thing is. The brittle ones say how a thing must be made, and then the ways of making change.
- They are written for someone reading in a hurry. Plain language, worked examples, and an honest section on what the agreement deliberately leaves out.
- They allow a graceful exit. Anything built on them can be read, moved and rebuilt elsewhere, which is why anybody trusts them in the first place.
What follows from that is a particular kind of progress, and it is very easy to miss. It does not arrive as an event. It arrives as a gradual fall in the cost of doing something ordinary, until one day the thing that used to take a specialist and a fortnight takes an afternoon and a short note. Nobody marks the moment, because there is no moment to mark.
There is a human element too. Agreements of this kind get written by people who are willing to sit through a great deal of patient discussion in order to produce a document that will never carry their names in public. That is a generous way to spend professional time, and the generosity compounds: the work holds up long after the circumstances that produced it have been forgotten.
The practical version, for anyone choosing between tools, is simply to ask how easy it would be to leave, and what shape the work would be in on the way out. If the honest answer is that everything made with it can be read and moved by something else, the choice is already a good deal safer than it looked.
Nicole Junkermann keeps the working definitions in the glossary and the longer sources on the reading list. The standards question in its applied form is set out in what responsible AI looks like when it is done well, and the companion note on why small teams ship clearly went up earlier in the week.