Share with your CTO
Your AI coding productivity metrics are probably lying to you. A senior engineering leader writing for HackerNoon describes how a team celebrating 40% more code per sprint traced a production incident to AI-assisted pull requests that each independently bypassed a centralized exception handler, each passing review, collectively destroying error visibility. The cleanup consumed a full sprint, and the productivity dashboard counted both the generation and the rework as output. GitClear’s 2026 analysis of 623 million code changes finds refactoring activity has collapsed from 21% of all changes in 2022 to 3.8% today, while duplicated code blocks are up 81%.
What this means for your business
The lines-of-code trap was supposed to be a solved problem. Fred Brooks buried it in 1975. What’s happened is that AI tooling rebuilt it with better branding. Merged pull requests and agent tasks completed carry the same flaw: they count artifacts of work, not the value the work delivers. When your engineering org grows its codebase 30 to 40% in a quarter without a corresponding growth in functionality, complexity grows with it, and complexity is what determines how fast the team can move in six months.
Microsoft’s study of Claude Code and GitHub Copilot CLI across tens of thousands of engineers found a 24% increase in merged pull requests, the largest agentic coding measurement published to date. The authors concede inside the paper that a merged PR is not the same as delivered value. That concession is doing a lot of work. Twenty-four percent more merged PRs is equally consistent with the same work arriving in smaller slices, which is exactly the behavior Uber’s diff-count dashboard produced before engineering got worse while the metric went up.
The signal worth watching: DORA’s 2026 ROI analysis finds that returns on AI coding tools track the strength of the surrounding engineering system, specifically code review capacity, test suite trust, and deployment pipeline tightness. Where those are weak, faster generation fills a queue and sits there. Before the next board presentation on AI productivity, audit whether your test suite is trusted enough that engineers actually act on a red build. Almost nobody does that audit. It’s the number that determines whether your AI investment is compounding or corroding.
Concept deep-dive: Error masking
Error masking happens when code catches an exception locally and either swallows it silently or logs it in a format the observability stack doesn’t recognize. It exists because catching errors close to their source feels safe and explicit. The problem is that centralized exception handlers exist precisely to normalize error shapes and trigger alerts. Inline handlers that bypass them are like individual store managers taking cash out of the register before it hits the accounting system: each transaction looks fine locally, and the books are wrong at the end of the day. GitClear finds error-masking constructs up 47% since 2023.
Based on reporting from AI Coding Productivity Is Not the Same as Engineering Progress, originally published 2026-09-16 23:49:00.
