The record-keeping side of this has a date attached to it in EU law. Article 26(6) of Regulation (EU) 2024/1689 requires deployers of high-risk AI systems to keep the logs the system generates automatically, to the extent those logs are under their control, for a period appropriate to the intended purpose and at least six months. Under Article 113(c) that duty starts 2 December 2027 for Annex III systems and 2 August 2028 for Annex I. For a company with no AI staff the practical consequence is procurement rather than architecture: the retained record is whatever the vendor's system emits, so a single-sensor loop leaves six months of exactly one number to answer with.
The support-agent example is the clean version. The dirtier one from the operator side: a growth loop measured on views will find the cheapest route to views, and the chart stays accurate the whole time. The fix that actually works is boring - a second counter that can contradict the first, reviewed together, on a schedule. Views next to paid sales, every night, on the same page. The moment both numbers sit side by side, the loop cannot hide behind the one it was handed. "Who can say no" is the right design question; in a one-person shop the answer has to be another metric, because there is no other person.
great article for the most part, ofc deterministic isnt synonymous with correct
The loop is a graph…
The record-keeping side of this has a date attached to it in EU law. Article 26(6) of Regulation (EU) 2024/1689 requires deployers of high-risk AI systems to keep the logs the system generates automatically, to the extent those logs are under their control, for a period appropriate to the intended purpose and at least six months. Under Article 113(c) that duty starts 2 December 2027 for Annex III systems and 2 August 2028 for Annex I. For a company with no AI staff the practical consequence is procurement rather than architecture: the retained record is whatever the vendor's system emits, so a single-sensor loop leaves six months of exactly one number to answer with.
AI Slop.
Great article, but I wonder wy table of contents doesn't fit content 😅
Table of Contents
1. A Loop Is a Control System With One Sensor
2. A Graph Is Not More Agents, It Is More Sensors
3. Four Ways a Single Loop Breaks and the Edge That Answers Each
4. The Real Design Work Is Deciding Who Can Say No
5. Graphs Fail in a Way of Their Own
6. Build the Loop First, Then Give It Neighbors
Thanks for writing this up. The graph shift is really about coded edges, so the model cannot optimize around the route the system already knows.
I turned these concepts into a runnable tutorial using Python, now with 800+ GitHub stars: https://github.com/hardness1020/awesome-agent-architecture/tree/main/sections/22-graph-engineering
The support-agent example is the clean version. The dirtier one from the operator side: a growth loop measured on views will find the cheapest route to views, and the chart stays accurate the whole time. The fix that actually works is boring - a second counter that can contradict the first, reviewed together, on a schedule. Views next to paid sales, every night, on the same page. The moment both numbers sit side by side, the loop cannot hide behind the one it was handed. "Who can say no" is the right design question; in a one-person shop the answer has to be another metric, because there is no other person.