Latest / Big Iron Radio / Homelab Brief - Safe Agent Writes: Approval Gates and Action Ledgers
Transcript
- 0:02Welcome to the Homelab Brief from Big Iron.
- 0:06Today, how to let an AI agent make changes to your Homelab
- 0:11without letting it run wild the right path pattern.
- 0:14Quick note, this is an AI voice drawn from a Big Iron guide.
- 0:18Links below.
- 0:20The mental model.
- 0:22Treat a right capable agent like a new hire
- 0:24you don't fully trust yet.
- 0:26You don't hand them root on day one.
- 0:29You make them propose, you review, then you apply.
- 0:32Three stages, never one.
- 0:34Concretely, every mutating tool should default
- 0:38to a dry run and return a diff.
- 0:40Here's exactly what I would change.
- 0:42A separate, explicitly authorized code path
- 0:45is the only thing that performs the real apply.
- 0:48The agent proposes.
- 0:50A different door commits.
- 0:52Now the subtle bug that breaks naive versions of this.
- 0:56Bind the approval token to a hash of the specific diff content,
- 0:59not to a call ID.
- 1:01Otherwise, the agent can regenerate a different plan
- 1:04and ride in on your stale approval.
- 1:07You approved that change the exact bytes of it not,
- 1:10whatever the agent does next.
- 1:12Make the human step livable.
- 1:15Route approvals through something like self-hosted NTFY,
- 1:18with action buttons approve or deny from your phone,
- 1:21and keep the token time to live sane, 15 to 30 minutes.
- 1:26Too long invites bypass.
- 1:28Too short trains you to rubber stamp.
- 1:30Both failure modes are real.
- 1:33And keep the receipts.
- 1:35In append-only, off-box, filesystem-hardened action ledger
- 1:39that links every propose, approve, and apply event by diff hash.
- 1:43When something goes wrong, that ledger is your only real post-mortem tool.
- 1:47Pair it with idempotent apply functions so a retry can't double apply.
- 1:52The full pattern's in the show notes.
- 1:55That's your brief.