Developer 2-Month Field Report: AI Agents Excel on Human-Written Codebases, Degrade Sharply on Greenfield and Agent-Modified Code

X/Twitter·medium signal

A developer with two months of intensive AI agent use published a conclusion that cuts through the hype: agents work well when adding features to existing human-written codebases, but performance degrades significantly for greenfield projects or codebases already heavily modified by agents. With 5,701 likes and 475K views, this is the most-engaged practitioner take in the dataset and directly corroborates the Alibaba long-horizon maintenance failure finding — human-written context appears to be a prerequisite for agent code quality, not an incidental factor. The implication: teams considering full agent-led development may be betting against the direction the evidence is pointing.