Glue code is the job

Every day, more than 1,000 student appointments depend on code I wrote to make two scheduling systems agree. Engineers file that kind of work under grunt work, and this winter's "AI software engineer" pitches leave it out of the demo. Both have it backwards. Glue is the job.

Since June, at a university IT department, I have connected those two systems, Starfish and iAdvise, through REST APIs in Python and JavaScript. I also wrote C++ controllers with the Zoom Rooms Controller SDK that cut the setup of a room from about 15 minutes to about 10 seconds for more than 200 users, and I built bi-directional APIs so that third-party applications can work with the rooms. None of it will ever be in a keynote.

The belief they share is the greenfield fantasy. It says real engineering happens inside one system you control, in a new service or a fresh repo with nothing in it to argue with. Connecting the old systems is plumbing, and plumbing goes to interns, to contractors and, as of last month, to bots.

On December 10, Cognition made Devin generally available at $500 a month, sold as an AI software engineer. It shipped with a Slack integration, an IDE extension and an API, so somebody at Cognition wrote a lot of glue. The onboarding advice is more modest than the landing page. It tells teams to start Devin on small frontend bugs, edge cases, first-draft PRs for backlog tasks and targeted refactors. Give Cognition credit for the honesty. That list is the greenfield fantasy's job description, and not one item on it asks Devin to make two systems agree.

Then Mark Zuckerberg told Joe Rogan this month that Meta will probably have, this year, "an AI that can effectively be a sort of midlevel engineer that you have at your company that can write code."

That is the cheap part.

I say that with a chat model open in a side window. Any chat model will write a function that maps one JSON shape onto another faster than I can type the field names, and a decent one.

But the mapping was never the hard part. Any two systems that store the same appointment hold two opinions about it. What counts as a cancellation. Whose clock the time slot is in. Which copy wins when both change in the same minute. Whether a deleted record is gone or only hidden. Neither system will change its mind for you. The day one side accepts a change that the other side rejects, you have two truths about one student, and the code in the middle has to decide which truth to tell. No ticket describes that decision.

The "AI software engineer" demos are cooking shows. The raw chicken goes into one oven, and a golden one comes out of another.

The most useful AI release of 2024 was a boring protocol for glue. Anthropic open-sourced the Model Context Protocol on November 25 because every assistant needed a custom connector to every tool, M times N of them, and it launched with servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer. Every server on that list is somebody's glue, and somebody will fix it every time the API on the other side changes. A protocol standardizes the plug. It does not tell you what a field means.

The greenfield fantasy is older than AI. Engineers sold it to each other long before any vendor did, because it flatters our vanity. New code gets a launch post and a line on a resume. Glue gets noticed once, on the day it breaks, and the rest of the time it belongs to whoever had the least standing to refuse it.

A conference room has the least forgiving acceptance test I know. The meeting starts on time, with the room or without it.

Stop calling it plumbing. Give the glue to your best engineer.