Software features evolve as developers extend existing code to meet new needs. Current AI can handle single requests well, but can they maintain correctness across multiple steps where each step builds on their own earlier code?
Researchers created FeatureArc, a benchmark that simulates this multi-stage evolution. It constructs tasks from real software repository histories, with each stage requiring the agent to extend its own code without any feedback about whether the stage passed. Tests verify that new behavior works and old behavior still passes.
Testing ten coding agents revealed a sharp drop: success fell from 62.7% on the first stage to 28.5% by the final stage. Agents frequently inherited errors from earlier stages and almost never detected them on their own. This shows that autonomous feature evolution without human checks remains a major challenge for current agents.