Former ByteDance engineer: Chinese AI practitioners generally rely on American agent tools, while domestic coding agents are "completely impractical"

According to Beating monitoring, Zhang Chi said in an interview that he uses Claude Code and GitHub Copilot’s Agent mode in his day-to-day work. His students use Claude Code for training and monitoring: they automatically check GPU status, and if an error is found, the system automatically debugs and restarts it; they also conduct inspections every 30 minutes—“then we can go to sleep.”

When asked whether all AI practitioners in China are using Claude Code, he first said, “Not all.” He said he personally is more accustomed to using Copilot, but then he confirmed that AI practitioners in China generally use American Agent tools. He said plainly that coding Agents for Chinese models are “completely impractical,” and noted that this is based on his personal experience rather than a system-wide evaluation. ByteDance’s overseas team can directly use Cursor, leading him to speculate that ByteDance’s future large models may be partly built with the help of American Agent tools. Zhang Chi also talked about a personal dilemma: Claude Code is so useful that he’s considering “whether he still needs to train PhD students,” but he’s also worried that if he doesn’t train the next generation of researchers, there will be no one left to do research in the future.