AI Agents and the Death of Human Control

Show notes

On this episode of Admin Access, Dominic and Evans dive deep into the growing chaos behind agentic AI workflows, and this time, it's not theory. Evans shares his hands-on journey with Openclaw, running on a dedicated Mac mini, and how it has fundamentally changed the way he works, thinks, and plans. From competitive intelligence and meeting prep to podcast research, this is a 24/7 digital employee running quietly in the background. But the wins come with warning signs. From an Amazon Kiro agent that took down an entire environment for 13 hours, to a Replit AI that deleted a whole production database despite explicit instructions not to, the incidents are stacking up fast. They break down why. An 85% per-step success rate compounds to just 20% across a 10-step workflow. Engineers are being mandated by leadership to adopt AI, with KPIs tied to token usage, while being expected to supervise work they can't fully review. And who owns the liability when agents go haywire? The engineer? The organization? The AI itself? Neither host has a clean answer, and they don't pretend to. Drawing parallels to Kenya's digital lending crisis and the cloud infra hype of 2012, Dominic and Evans make one thing clear. The gin is out of the bottle. The only question now is how you manage the blast radius.

AI Agents and the Death of Human Control

Show notes

In this episode