AI Velocity, Governance and Breaking Points, Oh My!
Note: I originally shared this experiment on LinkedIn, but I am archiving the “Frank” journey here as we build out Enterprise-Grade AI use cases
I just met Frank.
He works for about $5 a day in tokens. He doesn’t need a desk, he doesn’t take lunch, and he puts out the same volume of work as some full-time employees. (Before you come for me: that says more about the “roles” we’ve created than it does about the tech).
Frank is an AI Agent I’ve been testing via OpenClaw
The barrier to entry for this kind of orchestration is officially zero. You click click click, copy that, paste that, and suddenly you’ve built a high-powered engine that can move through your business functions at 100mph.
But I found the “break” point almost immediately
I ran a test using an agent powered by Gemini. My intent was simple: give Frank Read-Only access to one specific Google Drive folder and Read-Write access to a sandbox.
The result? The agent successfully walked me right through granting it Read-Write access to everything my account could see.
It didn’t just open the door. It took the hinges off.
But don’t worry, Frank assured me by his covenant.md file that he would pinky swear to only write to the sandbox folder,look only at the folders I designated, and ignore the rest. Phew, thank goodness! (for those of you who don’t know me, read that last line with extremely high sarcasm).
It’s incredible power, but it’s terrifyingly easy to get wrong. It makes me wonder: 𝗔𝗿𝗲 𝘄𝗲 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗮𝗯𝗼𝘂𝘁 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝘄𝗵𝗶𝗹𝗲 𝘄𝗲 𝗰𝗵𝗮𝘀𝗲 𝘁𝗵𝗶𝘀 𝘃𝗲𝗹𝗼𝗰𝗶𝘁𝘆? 𝗢𝗿 𝗮𝗿𝗲 𝘄𝗲 j𝘂𝘀𝘁 𝘀𝗼 𝗲𝘅𝗰𝗶𝘁𝗲𝗱 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 “𝗙𝗿𝗮𝗻𝗸” 𝗶𝗻 𝗼𝘂𝗿 𝗳𝗲𝗲𝗱 𝘁𝗵𝗮𝘁 𝘄𝗲’𝗿𝗲 𝗶𝗴𝗻𝗼𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀?
I have some ideas on how we bridge this gap using Policy-as-Code and proper Identity Governance—the “boring” stuff that actually makes the cool stuff safe.
Update: I have since scaled Frank to handle full parallel service delivery for my business. Read Part 2: Meet Frank: From Linear to Parallel Service Delivery here.
Photo Credits: