Your autonomous coding agent has shell access, so it can open sockets and touch files you never see. This share lays out the sandbox rules that keep every one of those moves visible, and keep the harness boring enough to trust.
Zero trust inside the sandbox, or do not run it at all. That is the whole test, and the setup is short.
What You Need
An autonomous coding agent that already has shell access.
Tutorial
Run each agent in an ephemeral container. Keep the harness dead simple, because the container is the sandbox and not a monitoring layer bolted on top of one.
Restrict the container's networking to loopback only. Nothing the agent talks to should need to leave the machine.
Write a rigid eval contract on the git diff. Judge every run by the diff it produces, since you cannot run agents on vibes.
Let the agent run the existing test suite inside the jail. It reports back only after that run, never before.
Wrapping Up
That is the whole harness: a throwaway container, loopback networking, a contract on the diff, and the test suite run inside the jail. None of it is clever, and that is the point, because boring local sandboxes beat complex monitoring layers every time. Set it up for one agent before you hand over your repo.
Ranked by how many people read each share in the last 7 days, each person counted once. Bots, the author and clicks from this list don't count. Shares from the last 14 days with a real write-up. Refreshed every 5 minutes.
Comments
Join the conversation — sign in to comment.
No comments yet — start the conversation!