Graph Engineering Explained: The Missing Fifth Layer of AI Agent Architecture
Every "my agent isn't working" postmortem starts the same way: someone rewrites the prompt. Adds a constraint. Adds an example. Ships it again. Three iterations later the agent still can't hold up in production, and the team is quietly out of ideas — because the prompt was never the layer that broke. There are five control layers standing between a raw model call and a system you can actually…
A postmortem for an AI agent that isn't performing well often begins with rewriting the prompt. This approach usually leads to three more iterations, after which the team may feel out of ideas because the real issue lies elsewhere. There are five control layers that stand between a raw model call and a reliable system capable of handling business outcomes: prompt, context, harness, loop, and graph.
Most teams only focus on the first two layers, leaving the remaining three unattended. The failures that occur in production, such as calling the wrong tool, repeatedly making the same mistake, or routing the output to the wrong reviewer, are primarily due to these overlooked layers. Graph engineering, the fifth and least understood layer, determines which component runs next, when agents work in parallel or sequentially, and where human intervention is required before any expensive or irreversible actions take place.
This article will break down all five layers, trace a single production failure from start to finish, and highlight the role of evaluation systems in each layer. The key concept is that there are five concentric layers surrounding the model, with each layer playing a crucial role in ensuring the system's reliability.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
