Research engineer says AI agents now write over 90% of his code

I use AI (Oct 2026)

A current research engineer at Salesforce AI Research reviewed his last 12 months of AI use and says machines now write more than 90% of his code. He runs about 5–6 coding agents such as Claude Code day to day, still reviews architecture himself and tests manually, and feels the fatigue of managing them in parallel. AI is his first stop for search and literature lookup, but complex research still needs judgment from human colleagues, and he keeps avoiding AI for writing his own personal expression.

Why it matters: The author turns recurring problems in generated code into review rules and concentrates human checks on architecture and test design, giving a concrete quality-control method for multi-agent work.