Software engineering got its safety rails decades ago: version control, blame, diffs, rollback, review. That machinery is why an engineer can accept help from an agent without fear — anything can be inspected, and anything can be undone. Knowledge workers have none of it. The analyst’s spreadsheet, the auditor’s working paper, the partner’s memo: no history, no attribution, no revert. Just files, and trust.
You cannot deploy an autonomous colleague into that world. Autonomy without undo is a liability with a login. So before Athena could be an employee, we had to build the equivalent of GitHub for people who will never learn git — and it turned out to be the hardest product problem we have taken on. Three reasons.
The unit of change is tiny and the volume is huge. An engineer’s commit is a considered bundle. A spreadsheet is ten thousand cells that change in bursts, mid-meeting, by three people and two agents at once. We version at the cell and block level, on top of CRDT sync — built with the creators of the Yjs ecosystem — so “who changed B4, when, from what” is a first-class query even while the document is being edited live. Attribution that arrives after a sync conflict resolution is attribution about the wrong world; it has to be woven into the merge itself.
Undo must cross the human–agent boundary. The undo people know is “undo my last keystroke.” The undo an institution needs is “unwind everything that procedure did across this workbook, without losing the edits my analyst made on top.” That is selective, attributed reversal — closer to revert-a-commit than ctrl-Z — exposed to someone who has never heard either term. Getting the semantics right took us longer than the engine.
The interface has to disappear. Engineers tolerate git’s vocabulary because the power is worth the learning curve. An audit partner will not learn “branch.” So the concepts wear plain clothes: history is a timeline you scrub, blame is “who did this,” rollback is a button on the version you trust. The test we hold ourselves to: a reviewer under deadline uses it without being taught.
Why it was worth it: this machinery is the documented reason our most conservative customer went the distance — every action attributed, every action reversible. Governance is not a policy document. It is a data structure.
— Product, New York