A Quick Lesson on Agents and Hallucinations
Most of my work is producing code. Getting things built is a large part of my responsibility. But I've been applying the lessons of Trust Topology wherever agents are involved, not just in software.
I had a chance to see what happens when I didn't have the right gate in place.
One of my flows is a content ingestion pipeline that pulls data from a variety of sources, categorizes it, then produces a structured marketing report. Some of the sources are websites. Some are complex enough that I have specialized agents for certain URLs.
Imagine my surprise when I was looking through a report and it referenced a company that doesn't exist. It completely hallucinated the name and the supporting data.
It would have been easy to blame the agent writing the report. But it's the structure, not the model, that provides trust in the output. So I dug in. The root cause was a mis-categorization early in the pipeline that cascaded through. An early agent decided one of the sources should be handled with the generic processor instead of the specialized one we built for it.
The result was a best-effort run through the funnel. The downstream agents followed their instructions perfectly, but garbage in meant garbage out.
The fix
I now have a deterministic verifier sitting early in the chain. A rerun of the pipeline succeeded. And I pushed the deterministic ceiling up just a little bit on my march towards perfection.
Read the full Trust Topology writeup →