A short piece of fiction, written by API Evangelist. The reader is imagined; the post, and every number and quotation from it, is real: "Okapi Is Dependabot For APIs, And The Business Needs The Register Behind It," Kin Lane, November 5, 2026, https://apievangelist.com/2026/11/05/okapi-is-dependabot-for-apis-and-the-business-needs-the-register-behind-it/
Dalia read the post on her phone in the car park, before going in. She runs the program office for AI work at a mid-sized insurer, and she had never opened a pull request in her life. She knew that because the post said people like her hadn't, and she bristled at it for about a second before conceding it was true.
The opening was easy to skim. Okapi calls itself "Dependabot for APIs", which meant nothing to her, and the paragraphs about extractors and branches slid past. She slowed down at the number. Of 841 breaking changes, 645 were model retirements. Three quarters of the feed said the same thing: the model you built on stops working on a date. The post named OpenAI retiring GPT-5 on December 11, and Dalia checked her own notes. She had signed off on three agent projects this year. She could not have said which models any of them ran on.
That was the uncomfortable part, and she let herself sit with it. The post said each retirement is, to the person who approved the budget, a re-evaluation, a re-prompt, a regression test and a quiet failure. It called that "a line item, not a bug". She had always filed it under bugs, as something the engineers would sort out. Nobody had ever sent her a line.
She did not buy all of it. The section walking one Stripe change through eight views felt like a catalog showing off. She doubted anyone in her building would ask for a view by regulation and a view by estate, and the pricing paragraph, per watched provider per business unit, read to her like a vendor deciding what it wanted to sell. The author admitted he was thinking out loud. She noted that, and noted that thinking out loud was not evidence a budget holder would pay.
But the workflow paragraph stuck. A quote-to-cash flow fails at step four on the effective date, and the agent running it does not know why. The post said an agent with a stale step doesn't degrade, "it fails confidently". She thought about the claims-intake agent her team had demonstrated in the spring. It had been confident then too.
She did not feel like signing up for anything. The tool was a GitHub app, the check API came by email, and none of that was hers to adopt. The register the post described, with a named owner and days to impact, did not exist at her company, and she suspected nobody could produce even the first column.
In the lobby she opened a blank spreadsheet and typed one row per agent project, three in all. She made two column headings, one for the model each agent calls and one for who owns it. She left the second column empty on purpose. Then she sent her platform lead a two-line message asking him to fill in the first column by Friday, and said she would work out the second one herself.
It wasn't a register. It was a start, and she thought the post would have called it plumbing.