Humans can do stupid sh*t, too, sure. But wrong or not, human artefacts are a manifestation of their intent. Text an agent edited after being told to do this or that is only derivatively so.

To be able to lean more comfortably into AI generation of artefacts, I want agents to treat everything a human has touched with much more respect. I.e. it should think twice before changing a line of text or code when it is clear the last one to touch it was a human. In contrast, when it was an agent, the barrier to change should be much lower.

I am thinking here also about agent memory, where I sometimes like to carefully curate a paragraph here or there, only to see it thrown out on an agent’s next edit.

But I think the same way about code. Code should be shaped according to need, of course, and I expect movement. But when I want to re-assert control over a vibe-coded codebase, I want to be able to insert myself by, for example, writing a carefully crafted new part of a test suite. Or a central piece of code. And I want this to become a stable island, or bedrock.

We have the means for this. Diffs of text files can “tell the story” of individual parts. What that requires is giving agents their own (git) identity. In short, I try to work under a model that will help me establish provenance (without the cost of marking text up).