Nearly 7,000 coding sessions. That's the sample size behind what may be the most detailed public documentation of an AI coding tool degrading in real-time.
Stella Laurenzo, AMD's director of AI, filed a GitHub issue on April 2 documenting specific behavioral regressions in Claude Code observed since early March. The numbers are precise: the tool reads existing code three times less before making edits, rewrites complete files twice as often, and abandons tasks mid-session at a rate that was previously zero.
What "Reading Less" Costs
In AI coding tools, reading context before acting is the mechanism that prevents the tool from breaking things it doesn't understand. When an AI reads the function you're editing, it knows what variables are in scope, what the surrounding code expects, and what it shouldn't touch. When it skips that, it's editing with incomplete information.
A 3x drop in pre-edit code reading directly explains the second data point: whole-file rewrites happening twice as often. If you don't understand the specific function you need to modify, rewriting the whole file is a blunt workaround. For users, this means more time reviewing diffs, unexpected changes to code that didn't need touching, and regressions introduced by edits that looked fine on the surface.
The Abandonment Rate Is a Different Kind of Problem
Task abandonment moving from zero to a measurable positive rate is distinct from the other regressions. When a metric goes from zero to positive, a behavior that didn't exist before now does. Some threshold in how Claude Code evaluates "can I finish this?" shifted between February and March.
A coding assistant that stops mid-task is worse than one that finishes badly. A bad completion gives you something to correct. Abandonment means starting over.
Laurenzo's 7,000-session sample is large enough to separate signal from noise. This isn't frustrated users venting - it's a structured analysis from someone who builds AI systems professionally at AMD. Anthropic has not publicly addressed the findings since the issue was filed on April 2.
The timing matters. Claude Code has been updated at high frequency since its standalone launch in early 2025. That pace creates more surface area for unintended regressions. Whether the March changes were deliberate model adjustments or accidental side effects, the behavioral impact on user workflows is documented and measurable.
AI coding tools compete on reliability as much as capability. A tool that reads less, rewrites more, and gives up more often has a direct productivity cost that compounds across every session a developer runs. Teams that depend on Claude Code for production work are now weighing whether to wait for Anthropic to address what changed or start building verification steps into their own workflows.