Reconciling two large folder trees with read-only AI subagents
Two overlapping copies of years of operations files needed to become one, and the copy everyone assumed was safe to delete turned out to hold an entire subject folder with no counterpart anywhere else. Finding that meant refusing to trust the 91 same-name-different-size conflicts at face value and actually comparing content, not just file sizes.
Category
Read-only reconciliation with parallel agents.
What I was trying to accomplish
Find everything unique to a personal OneDrive copy of years of operations files, move it to the shared SharePoint source of truth, and identify what could safely be deleted, without losing a file.
The data challenges
Two large, only partly overlapping folder trees, 91 same-name-different-size conflicts that needed real content comparison, not just byte counts, to explain, and instructions that changed mid-run.
The result
A reconciliation report, a copy list, a safe-to-delete list, a destination-only list, a folder to copy whole, and a method-and-caveats tab, executed only after review, plus one explicit catch: an entire subject folder with no real counterpart at the destination, which changed the headline from clean this up to do not delete the source folder yet.
Technical details
Tools: Claude Cowork, parallel subagents, Microsoft 365 (OneDrive and SharePoint), Excel
Years of operations files existed in two places: a personal OneDrive folder and the shared SharePoint site that was supposed to be the single source of truth. The two trees overlapped only partly.
Claude dispatched several subagents in parallel to inventory matching branches of both trees. Each was explicitly barred from any write, copy, or delete action and wrote its inventory to disk. Claude then built a spreadsheet that classified every item, holding same-name, different-size files as conflicts until they could be explained. There were 91 such pairs. Rather than trust the byte counts, Claude compared the extracted text of a sample of pairs across every Office format, found the differences were document-library metadata rather than content, and said plainly that this was strong evidence, not proof.
One subagent returned a summarized listing instead of a complete one, which produced an unreliable comparison, so Claude caught it and sent a second read-only agent to redo that branch. When manual changes happened mid-task, Claude recomputed the baseline and pushed back with data where an assumption about what had changed was wrong. Deletion of the source folder was deferred until a large sync finished and a file-by-file check could confirm nothing was missing.
Pull up a chair.
Let's talk through where things stand, and where they could go.