LangGraph

A dictated prompt becomes an episode — if one agent resolves facts, a second tightens prose, and a human approves the dispatch.

Some models read text, score it, and return a number. No chat, no reasoning, just decisions — and they're running under every router you use.

Most agent search runs add no new evidence — and 66-85% of the context window is just search snippets. Here's what actually stops the loop.