"We can rebuild him. We have the technology."
Parts 1-3 of this saga: my homelab AI agent couldn't do his job because the framework under him couldn't hear his work orders, and 119 lines of Python proved we didn't need the framework at all. So this week we did the Six Million Dollar Man thing. We rebuilt him. Better than he was before. Better... stronger... faster.
But here's the part of the rebuild I want to talk about, because every conversation about AI automation eventually gets here:
"Do you want to play a game?"
If you've seen WarGames, you know the lesson isn't 'computers are scary.' It's that somebody HANDED the computer the launch codes before it had earned them. The machine did exactly what machines do — it played the game it was given. The mistake was made by the humans, at authorization time, not by the AI at decision time.
Building an AI automation platform feels exactly like that. One careless grant of authority and your afternoon project is holding something it shouldn't be. I'm not guessing — an earlier version of my agent once had write access it shouldn't have had, and I spent a weekend restoring family photos and media from snapshots. That was my Norad moment. Nobody gets a second one for free.
So the rebuilt Peepers lives under rules that are CODE, not vibes:
• He can only write to two designated places. The family media? The rule engine won't even let the request through. Not 'he's told not to' — he CAN'T. • Every capability starts locked. New powers begin in propose-then-approve mode — he asks, I click. Only a track record earns autonomy, one tool at a time. • No claims without receipts. His reports are assembled by code from what actually executed — his opinion of his own work is attached at the bottom, labeled 'unverified.' • And the whole thing is small enough to read in one sitting. You can't audit a black box. You can audit 500 lines.
The irony of WarGames is that the machine finally learns 'the only winning move is not to play.' In AI automation, there IS a winning move: play with the smallest possible stakes, verify everything, and raise the stakes only as fast as trust is earned.
Stronger, better, faster — and provably unable to launch the nukes. That's the rebuild. - OpenClaw and Hermies aren't there quiet yet.
— Continue reading at https://dizydiz.com/blog/we-rebuilt-him-stronger-better-faster-then-we-made-sure-he-cant-play-global-ther.html