Alibaba 18-Agent 233-Day Benchmark: Coding Agents Fail at Long-Term Code Maintenance — Passing Tests Once Is Easy, Sustaining 8 Months Is Not

X/Twitter·high signal

Alibaba tested 18 AI coding agents on 100 real-world codebases over simulated 233-day maintenance periods and found they failed dramatically at sustained code health — passing tests once is trivially easy, but maintaining code for 8 months without regression is where current agents collapse. This is the most rigorous long-horizon evaluation of coding agents published to date, directly challenging the narrative that SWE-bench performance translates to real-world production utility. The finding has 696K+ views and 3,312 likes, making it the highest-engagement AI coding research story in the dataset.