Agent Graph Engineering, Part 0: I Needed This Before It Had a Name

Description and use

<blockquote> <p><strong>TL;DR</strong> Building a chatbot is easy. Building an AI system you can trace, test, and change without fear is not. I spent years shipping LLM features as intent detection plus nested <code>if/else</code>, and no amount of observability tooling fixed it, because the problem was that the flow only existed in code, in my head, and in a stale diagram. The fix was to stop writing the flow as control flow and start declaring it as a graph. This post is the story of how I got there, years before the practice had a name, and why I think graphs are a durable answer rather tha

Problem

<blockquote> <p><strong>TL;DR</strong> Building a chatbot is easy. Building an AI system you can trace, test, and change without fear is not. I spent years shipping LLM features as intent detection plus nested <code>if/else</code>, and no amount of observability tooling fixed it, because the problem was that the flow only existed in code, in my head, and in a stale diagram. The fix was to stop writing the flow as control flow and start declaring it as a graph. This post is the story of how I got there, years before the practice had a name, and why I think graphs are a durable answer rather tha

Outcome

<blockquote> <p><strong>TL;DR</strong> Building a chatbot is easy. Building an AI system you can trace, test, and change without fear is not. I spent years shipping LLM features as intent detection plus nested <code>if/else</code>, and no amount of observability tooling fixed it, because the problem was that the flow only existed in code, in my head, and in a stale diagram. The fix was to stop writing the flow as control flow and start declaring it as a graph. This post is the story of how I got there, years before the practice had a name, and why I think graphs are a durable answer rather tha

Evidence

Agent Graph Engineering, Part 0: I Needed This Before It Had a Name

Source