I agree that the harness is where a lot of the practical economics lives, but I’d separate orchestration from one quieter cost: how much already-seen state gets serialized into every next call. A model swap can hide that variable, whereas a deterministic local compression step gives the harness a reproducible place to measure it before adding routing or memory complexity. One small tool I’ve been evaluating is Contextpress (GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub · contextpress · PyPI); it operates locally on the message list rather than making a second model call. I’d compare harnesses on cost per successful task, retries, and time-to-result, because “more capable model” and “less repeated context” are different levers.
1 Like