Your Agent Doesn't Need an LLM for Half Its Decisions—Use a Classifier

Arindam_1729 · x · 2026-10-05

The author argues half of an agent's decisions don't need an LLM: "Is this urgent?", "Should I retry?", "Which tool to call?", "Is this safe?", "Category A or B?" are routinely sent to the same expensive model that writes your code and spends 2,000 tokens thinking—costly for a yes/no.

A three-part test for whether a decision needs reasoning: can all valid answers be defined beforehand, would a human settle it in seconds via pattern recognition, and does it happen thousands of times? If yes to all, you need a classifier, which changes the agent architecture considerably.

Original post →

Your Agent Doesn't Need an LLM for Half Its Decisions—Use a Classifier

More from coding & agent