Architecture where Jev judges and the LLM executes

Dev Agrawal @devagrawal09Originally posted 2026.09.20

Open original on X

People are too focused on getting their LLM agents to orchestrate Jev calls It should be the other way around The bulk of the work should be handled by deterministic workflows using Jev for judgement, including deciding when to delegate to an LLM and what context to provide it

This post quotes another post: “Stanley: a Jev-first self-evolving coding agent”