What it covers. Write access covers any capability that writes to the file system — generating a report, editing source code, saving a download, updating a local database. It's narrower than shell access (no arbitrary commands) but still a real disk-modifying permission.
Scope matters. Write access to a single output directory is qualitatively different from write access anywhere. A capability that promises to touch only ./output/ still needs to be honest about it, so AgentMaps records write scope on the detail page when the source declares one. Treat unbounded write scope the same way you'd treat shell access.
When to be careful. Anywhere the write path is under the AI's control — "save this to the filename I say" — treat it like limited shell. Prefer capabilities that write to a well-known project directory over ones that ask for a home-directory-wide write scope. And review generated output before running it, especially when the capability writes executable files or configs.