A pure-LLM agent asks the model what to do at every step: expensive and never the same twice. A hybrid agent runs the work as deterministic code, reserves AI for real judgment, and improves the code between runs. Consistent and cheap at run time; smarter over time.
