Essays on building a business-facing AI management platform.
Published irregularly by the StoneForge engineering team.
Our two biggest dev token sucks, and why the fix is also the lock.
The CMS goes up this week. Here's the thinking it sits on, one idea at a time.
A StoneForge for Dummies guide to the layered Instruction Stack, plus the story of how we arrived at Skills eight months before anyone called it Skills.
Why state machines break the minute you put AI in a node, what cells do differently, and a workflow run that survived a model swap and a provider outage in the same Wednesday.
Why a transcript is not memory, what the four-channel pump captures, and the question a regulator actually asks when something goes wrong.
What each of the five layers does, what each one would have caught at Lily, and the six-question buyer's checklist for the next AI platform you sign for.
An essay on the structural primitive long-context LLMs are missing, what the McKinsey/Lily incident proves about the gap, and what the CRUD+ cell looks like in practice.
What changes when the decision cost collapses, why packet switching is still teaching us the same lesson, and why the AI-protocol standards wars may have already lost to Apple's accessibility API.
Fifty-year technologies, the AI-consumable data problem, and why the next common protocol will look more like chmod than like JSON.
How StoneForge began in a Louisiana bayou truck stop, and why the Request for Comments process remains the finest model of engineering humility ever devised.