Pulling yes-or-no judgments out of the LLM and handing them to Jev
After scanning his own agent loops, the author found many binary questions being sent to a frontier model built for writing, and lays out a five-step process for moving semantic branches to Jev.
anthropic, openai and everyone else keep dropping better models, nobody's checking how many of those calls even needed to happen went digging inside my own agent loops and found dozens of plain yes or no questions being handed to a model whose whole job is writing paragraphs jev engineering: stop letting a paragraph writer answer yes or no here's the pass that fixed it: ~ scan, surface every hidden judgment sitting between the visible actions ~ cost, measure each one in latency and tokens on its own, not averaged across the whole run ~ type, hand the semantic branches to jev instead of an llm ~ gate, classify risky tool calls before they execute, never after ~ verify, judge completion by evidence in the environment, not by the model just claiming it's done most agents route every single decision back through the same frontier model, twenty calls in and the agent has spent more time deciding whether to keep going than actually doing the work splitting generation from judgment from enforcement is the whole fix, not a better prompt, just a cleaner boundary between the three full architecture, the failure modes and the rollout order are in the breakdown below
This post quotes another post: x.com/cryptowluha/status/2103600863868604579