# Jev for agents.

These are ideas for Jev inside a coding agent. Most come from [a post by Diogo Almeida](https://x.com/completeskeptic/status/2101894250401271876). Diogo works at TypeSafe, the maker of Jev. Each cell names one decision and quotes the question Jev would answer.

Permission: "Should this command run?" If Jev is not sure, the agent would ask a person. Tool choice: "Which tool fits this step?" The agent would load only the top few. Context: "Does this chunk matter now?" The agent would hide it, summarize it, or show it whole.

Model choice: "Is this step easy?" An easy step would go to a smaller model. Parallel work: "Can these tasks run at once?" The agent would split them. Instructions: "Is this front-end work?" The agent would load the style guide.

Data safety: "Could this touch secrets?" The task would run on an approved model. Done check: "Is the task finished?" The agent would stop or keep going. Review: "Does this change do what was asked?" The agent would approve it or send it back.

Evaluation: "Which prompt wins?" A language model asked to grade writes its grade in free text. A Jev judgment comes back as a probability, and you can check it against cases a person labeled. ThinkThen added the done check and the prompt comparison.

Each decision is one small question with a probability.

On GitHub: [github.com/botassembly/beatles-bench](https://github.com/botassembly/beatles-bench)