The Edit Was Supposed to Be Surgical. It Rewrote the Whole File.

My research agent, Hermes, has overwritten one of my working documents a few times now.

The pattern is the same every time. It pulls in a doc, decides to make a "surgical edit," and ends up writing back the entire file. What lands on disk is its version of the document, not mine.

The last time, the doc was tracking the status of a set of changes. I rebuilt what I could from emails, other chats and files on disk. Some of it I couldn't get back. It was painful.

Here's what I took from it. I had been treating "edit" as one action. It's two: add something, or replace everything. The agent thought it was doing the first and the tool did the second. "Surgical" is a feeling the agent has. The tool can't tell surgery from a rewrite.

When Hermes later audited my memory stack, its list of requirements included one that now reads like a postmortem: a safe, append-only way to save notes, because the write path that overwrites has real data-loss failure modes. Before you let an agent write automatically, give it one narrow verb that can only add.

I have a standing append-only skill now. It doesn't always work. Sometimes it takes more than one pass to get it right, and I won't tell you this one is solved.

A skill is an instruction. Instructions get misread. A tool that physically can't overwrite can't misread anything.

Narrow the verb before you scale the volume.

What's the most destructive thing your agent's edit tool can do, and is it also the easiest thing for it to reach for?