Post

I think platforms like github should invest in exposing audit information. An agent should be able to see a log of events that occurred and know whether it was me, another agent, or itself that made some change.

We require all LLM PRs opened in draft mode at first. Very often I'll mark the PR as ready for review but then I'll fix something flagged in review. The LLM will get confused that the PR is no longer a draft and it'll mark it as draft again...

The LLM should know that it was an intentional decision and not a mistake it made. I'm working on an experiment that will give each session their own access tokens, but fundamentally it'll require effort from the services to store some identifiers.