MDK 0.11 Ships an AI Agent That Still Asks Permission Before Touching a Rig

The reference agent inside MDK 0.11 doesn't just act. It stops. It pushes the request up the chain and waits for a human to approve it.

That's the story under the version number. Site identity and command controls now decide who may issue which instruction, and the agent can't execute until someone with the right credentials clears it. Version 0.11, not 1.0. The leash is built in on purpose.

Paolo Ardoino has been loud about the endgame. He wants AI running Bitcoin mines. Fewer bodies on site. Machines adjusting hash rate in real time, spotting a failing fan before it turns into a fire, and selling power back to the grid when the spot price beats the block reward.

Good pitch. And still mostly a pitch.

Here's the thing people keep missing. Autonomous mining isn't an intelligence problem. It's a permissions problem. A wrong command at a 100 megawatt site doesn't cost you a bad tweet. It costs you a transformer, a rack of S19s, or an insurance claim you lose. The invalidation point sits at the first operator who disables the approval flag because it's annoying and slows down his night shift.

So the interesting decision in 0.11 isn't the agent. It's the identity layer. Who's allowed to trigger a shutdown. Who's allowed to push a firmware change across 4,000 machines. Get that wrong and you've automated your worst day.

Historically speaking, every industrial automation cycle has run the same play: sensors first, then remote control, then autonomy, then a hard walk back toward human checkpoints once the first expensive accident lands. Bitcoin mining is running that play at speed, and MDK is trying to skip ahead to the checkpoint.

The winners here are the mid-size operators. A shop running three sites can't afford a 24/7 NOC. A shop running thirty already has one and doesn't need the approval gate. The build-it-yourself crowd gets the most upside, and the most rope.

My read: human-in-the-loop is the correct default, and the teams that treat that gate as a feature rather than friction will be the ones still hashing after the first ugly incident. Watch whether the approval step survives contact with a real profit-and-loss sheet.