Why I prefer agents with explicit state

When an agent keeps its “state” only in the conversation history (what was said, in what order), debugging an unexpected behaviour means re-reading the whole exchange to reconstruct what the agent “believes” at a given point.

With explicit state — a structured object (confidence, documents_found, current_step…) that every node in the graph reads and updates — the same debugging becomes: print the state at step N. That’s the difference between re-reading a novel to understand a character, and reading their file.