Jev in production › Model and agent routing

JevAgentHarness

Jev routes each coding-agent step to a fast or frontier model, with vendor claims of savings while matching accuracy on 74 Terminal-Bench 2.0 tasks (JevAgentHarness).

Open on GitHub ↗

74measured against a baseline, as published by the source
Use
Model and agent routing
Industry
Developer tools
Form
Open-source tool
Stage
Beta
Listed
2026-10-16
Found via
github
Repository
JevAgentHarness/JevAgentHarness
Stars
60
Forks
12
Last push
2026-10-16
Language
Python
License
MIT
JevAgentHarness screenshot
assets/results-light.png in the JevAgentHarness/JevAgentHarness README, MIT; shown from GitHub.

The README opens with

A small decision model watches every step of your coding agent. Routine steps go to a cheap model, hard steps to the strong one. Same results, a fraction of the bill.

Project page · Results · Quickstart · Harnesses · How it works · Decision points · Reproduce · Models and data · Paper (arXiv) · Citation

Coding agents call a frontier model at every step: to read a file, to run ls, to re-run a test. Most of those steps are routine and a fast model does them just as well. A few steps decide the outcome: planning the fix, debugging a failure, deciding that the task is done. JevAgentHarness keeps those on the strong model and moves the rest to a fast one. It is a local OpenAI-compatible proxy, so the agent itself does not change.

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for model and agent routing