Overview & result
zjunlp shows JevLoop, built on Jev: An agent loop where the decisions cost no LLM call.
From the original post
JevLoop by @zjunlp: An agent loop where the decisions cost no LLM call.
How Jev fits in the loop
- The agent loop is declared as decision nodes in a DECISION.md file; the reference loop has seven.
- Each fork (act or not, which tool, which file, is it safe, did it work, is it done) goes to a decision model as a pick, a score or a yes/no.
- The backend is the hosted Jev API if a key is set, otherwise a local Laya server.
- The LLM is called only to write: once, plus at most one revision.
How to reproduce
- Review the source documentation and repository at https://github.com/zjunlp/JevLoop.
- Configure environment with TypeSafe Jev / System One API keys.
- Run test prompts against the decision endpoints.
Why this build matters
Shows Jev in practical use for browser and autonomous agents.
Limitations
- The quick-start run uses a rule table as judge; the README says it shows the shape of the loop, not decision quality.
- With no backend available, every step escalates.
- The README's own numbers show decisions can take a large share of wall-clock time on a local backend.