One negative is that Claude Code is pretty buggy, and Anthropic makes frequent changes that cause unexpected regressions [0]. With the harness now doing weird stuff with proxies, I'd be worried of them inadvertently introducing bugs which affect people using the feature legitimately.
[0] A recent example: https://www.anthropic.com/engineering/april-23-postmortem
Maybe they should try running Mythos to check Claude Code, given their marketing with it's superior performance.