A little humor to start your week! I found this way too relatable, mirroring how I feel sometimes when I try to talk about AI with people I meet or know who don't use the tools and can't see beyond the negative narratives. I don't judge it, it's biology, and a repeated pattern that comes with every big tech shift or invention. They'll get there in their own time. Grateful for finding my fellows Straxs though! Have a great week everyone! (S7E6 - The Snowmen)

Most of my agent work is reversible: it writes a file, I read it, I fix it. The runs that keep me up are the irreversible ones — the send, the publish, the payment, the delete. Once it fires, there's no diff to review. The pattern I've settled into is draft-for-approval: the agent does everything up to the irreversible step, stops, and hands me an exact preview of what it's about to do. I approve the specific action, not the general behavior. It's the Cynefin split in practice — reversible work runs unattended, irreversible work puts a human on the gate. But "human on the gate" has a failure mode: approval fatigue. Approve enough previews and you start rubber-stamping, and a rubber-stamp is just a slower auto-approve. So two questions I'm chewing on: how do you decide which actions are irreversible enough to gate vs. reversible enough to let run — a fixed list, or something the workflow classifies? And how do you keep the gate meaningful instead of a reflex click — batching, a forced diff, a cooling-off step? Curious how others draw the line and keep the checkpoint honest.