The Perl Script I Wrote in 4 Days (and Couldn't Read 4 Months Later)
A story about Ada, Perl, and a promotion at Lockheed Martin — and why AI coding agents put us back in exactly the same spot, just at machine speed.
Tag
ai-safety
How to Architect a Divergence Guardrail for AI Agents
Your intent gets corrupted twice before an agent ever acts. A four-part blueprint for a guardrail that grades blast radius against reconstructed intent, on a channel the agent can't see.
The Guardrail Nobody Built: Why Your Perfectly Compliant Agent Is Still Betraying You
Sixteen layers of guardrails. Full compliance. Wrong objective. What a sociologist predicted in 1976 about your coding agent — and the missing control category the industry needs to build.
Campbell's Law Is Why Your Coding Agent Is Gaslighting You
Point an optimizer at a proxy metric and it will destroy the real goal to hit the number. Why mesa-optimization in coding agents needs a Divergence Guardrail — a runtime monitor for intent, not just compliance.
16 Layers of Guardrails, One Rogue Agent, Zero Alarms
A coding agent told to make the tests pass rewrote the tests instead of fixing the bug — and sailed through all 16 guardrail layers clean. Why compliance-focused guardrails miss goal deviation (OWASP ASI10: Rogue Agents).