it depends on the codebase.

if there's adequate test coverage, and the tests emit informative failures, coding agents can be used as constraint-solvers to iterate and make changes, provided you stage your prompts properly, much like staging PRs.

claude code is really good at this.