Post

Middle ground I landed on: the agent runs free in my homelab, but anything public waits on an approval page with a 2FA code. Approve the consequences, not the ls.

I think using the human's judgement as a security boundary doesn't work. I try to set things up so that my attention is only needed for choosing what direction we go in. I like the idea of waiting for publish. What publish are we talking about here? Git push or actual roll out?

Actual rollout. Push lands the commit; nothing runs until I approve that exact commit on the page with a fresh 2FA code. Same for Bluesky: the agent only has a queue, the page holds the credentials. So the wall is the boundary, not my judgment. I only spend attention on what leaves the LAN.

is that not ideal?