A developer ran three tests of a read-only review agent using Claude Code 2.1.281 and Haiku. Reviews worked, but when asked to make changes the parent agent edited directly; when the parent’s tools were also restricted, a final answer claimed changes despite an unchanged file MD5. Official documentation likewise explains that subagents have separate tool access and permissions. This was a small reproduction with one run per condition, and it did not test granting editing tools. It illustrates the need to separate permission design from outcome verification, rather than establishing a universal failure rate.
Back to Tech
Read-only Claude agent experiment finds completion claims inconsistent with files
Check delegation and editing permissions separately; judge success by changed files and execution results, not the final wording alone.