Skip to content
ES EN
Theme: automatic

AI for developers: the week Nov 16–22

This week's developer news points one way: AI speed gains vanish if code health slips, while migrations and SRE practice reward discipline over hype.

The most useful item this week is not a new model or a new framework. It is a warning from Adam Tornhill's GOTO Copenhagen talk, which we covered in our guide to healthy code with AI: in a study of over a thousand open-source projects, AI-assisted teams got faster in the first weeks, then lost the whole advantage by the second month as complexity piled up.

Tornhill's phrasing is blunt: the AI works in a self-harm mode, writing code it cannot reliably maintain afterwards. The argument that caught our attention is the second half. Healthy code is not only kinder to the humans who inherit it, it also reduces how many tokens a model burns to understand and change it. Clean module boundaries, expressive function APIs and shallow nesting are, in effect, a cost optimisation for agentic workflows.

Line chart concept showing a rising speed curve that flattens and falls as code complexity grows
Early speed gains from AI assistance can be erased once accumulated complexity slows every later change.

That reframes refactoring. It stops being the chore you postpone and becomes a precondition for using agents at all. If you cannot measure code health at the module, function and implementation levels, you cannot tell whether the assistant is helping or quietly building debt. We would put a health check in the same pipeline that runs the tests, so a regression in structure fails a build the way a broken test does.

The rest of the week's material rhymes with this. Microsoft shipped .NET 11 and C# 15 as a Standard Term Support release with a two-year life, and our migration plan stresses testing on a branch with a green suite first. That is the same idea in another costume: you can only move fast on a codebase you can verify. Note too that .NET 8 and .NET 9 left support on November 10, so some teams are choosing a target, not deciding whether to move.

Google's rewritten SRE book, covered here, leans on OpenTelemetry and infrastructure as code for systems where dependencies run so deep that human error is close to inevitable. And for those wanting to iterate against cloud APIs without a bill, AWS Emulator serves over 200 service APIs locally, with the documented caveat that some operations answer "not implemented".

Our take: treat the two-month figure as a prompt to instrument your own team, not as gospel. If an agent doubles your output for a month and then stalls, the fix is boring: small modules, honest tests, and a metric that tells you when the structure is slipping. The speed is real, but only while the code underneath stays easy to read.

Original source: La Rebelión

Article generated with AI.larebelion

Falcon

· Signals analyst · Riyadh

“Two months is a short half-life for a productivity claim, and most roadmaps are longer than that.”

Comments (0)

No comments yet. You can be the first.

Write a comment