Graham McNicoll

A year ago, most engineers wanted to see every line an AI wrote before it went near their codebase. They read the diffs. They were skeptical because they didn't trust it. Today, a lot of those same engineers have not written much code by hand in months. They describe what they want, the agent produces it, they trim the copious comments, and they ship. What I notice is how fast the human checks are disappearing. The review step that felt non-negotiable a year ago has quietly atrophied. Now we use AI to check the work of AI. This actually works out better than I would have predicted; I am not arguing for going back. The productivity improvements are impressive. But the real issue is that while code can work, and AI can say it's good, it can still have unexpected effects on your users and business. The code may quietly degrade something for your users, or create a strange UX, and at the speed product teams are moving these days, nobody is well positioned to catch this anymore. The agentic era is exactly when measurement should matter most. If a human is no longer standing between a change and your users, guardrail metrics have to be. Wrap the change behind a flag. Roll it back automatically when a guardrail metric degrades. Even feed back this degradation into the AI to self-improve and try again. How much of what your team shipped last month was actually read by a person before it went live? How many releases were tested to make sure there were no unintended side effects? We built safe rollouts in GrowthBook for exactly this; more at growthbook.io.