Original articles
Measure the second change when AI writes the first
DevMojos argues that an AI coding pilot should measure the effort to make a follow-up change as well as the time it took to ship the first one.
In an InfoQ interview following his talk at GOTO Copenhagen, Adam Tornhill described three things he uses when refactoring with AI agents. He establishes a baseline, protects the behaviour that must stay the same, and gives the agent feedback as it works. We’d bring that approach forward to the moment a team decides whether its AI coding pilot is paying off.
Imagine an agent adds a new returns rule to an internal support tool. The first ticket closes quickly and the demo looks right. A month later, the policy changes again. If that rule now lives in the support form, the service and a reporting job, the next developer has to find every copy and work out which one takes precedence. The effort from that second ticket belongs in the assessment of the first.
We’d give the pilot two tickets. For the first, make one real change and record the time from a clear brief to a reviewed, passing change. Then ask a different developer to make a plausible follow-up in the same part of the code. Keep the existing tests and ordinary review requirements. Record how long they spend tracing the behaviour, how much rework review finds and which files had to change.
If the follow-up is harder than it should be, fix the cause while the example is still fresh. Put the returns rule in one place and test it where the decision is made. Ask for another small variation and see whether the next change is easier to review. Token usage can be useful here when the tools expose it, alongside engineering time and the work needed to get a change through review.
The question we’d take to a team lead is whether the next person can change what the agent just built. A pilot that can answer that question gives you a firmer basis for expanding agent access across a codebase.
For background, see Ben Linders’s InfoQ interview with Adam Tornhill in “Producing Healthy Code with Artificial Intelligence Costs Fewer Tokens” (19 November 2026) at https://www.infoq.com/news/2026/11/healthy-code-AI-tokens/
This story was drafted with AI assistance.

