Letting agents run unsupervised: blast radius, sandboxes, and rollback

Agentic coding tools will happily run commands without asking. We're getting into how to let them, safely: sandboxes, blast radius, and getting your work back when it goes wrong.

FIRST TIME? This is a small, friendly group that meets twice a month at Schuler Books on 28th Street, in the Chapbook Cafe. Free, no cover, nothing to prepare. Turn up, grab a coffee, and talk shop.

WHAT WE'RE DIGGING INTO
When it's reasonable to let an agent run without approving every step, and when it absolutely isn't. Sandboxing and containers so a bad command can't reach anything that matters. Git hygiene that makes rollback boring instead of terrifying. The failure stories — what people have actually lost and how.

WHO THIS IS FOR
Anyone using Claude Code, Cursor, Codex, Copilot or similar. People who've typed the skip-permissions flag and felt a bit nervous about it. Also anyone who hasn't touched these tools yet and wants to see what the fuss is — beginners genuinely welcome, and questions are the useful part.

HOW THE EVENING GOES
6:30 - coffee, introductions, what everyone's working on
6:50 - I show my setup and what it's allowed to do unsupervised
7:20 - open floor: bring a project, a problem, or just listen
8:30 - whoever's still talking keeps talking

Bring a laptop if you want to follow along, or bring nothing. Schuler Books, 2660 28th St SE - big parking lot, easy to find. Look for the cowboy hat in the cafe.