TianyuCodings/JevHarness

LLM-authored task-specific Jev harnesses with optional full-trajectory reward reflection and GEPA evolution.

What it solves

JevHarness addresses the high latency and cost associated with using large, reasoning-heavy LLMs for every step of a task. It allows developers to offload repetitive, context-sensitive decisions to a faster, lightweight judgment engine (Jev) while keeping the complex reasoning logic in a structured, "frozen" harness.

How it works

An authoring LLM generates a task-specific harness—a combination of Python code, structured instructions, and Jev-specific criteria. This harness processes task observations into features and uses Jev to make fast, fuzzy decisions based on those features. Once the harness is developed and validated, it can be frozen and executed independently of the authoring LLM. The framework supports optional refinement through reflection, where execution traces and rewards are used to iteratively improve the harness.

Who it’s for

This project is for developers and AI engineers building agentic workflows who need to balance high-level reasoning with low-latency, reliable execution. It is particularly useful for tasks that require repeated, rapid decision-making within a defined environment, such as game-playing agents or automated routing systems.

Highlights

  • Hybrid Execution: Combines the deep reasoning of an authoring LLM with the speed of lightweight Jev judgment calls.
  • Iterative Optimization: Includes support for reflection and candidate selection (GEPA) to improve harness performance based on actual execution traces.
  • Claude Code Integration: Provides a plugin for Claude Code to help agents build, test, and refine harnesses directly within their development environment.
  • Transparent Debugging: Includes tools to inspect decision records, view evolution trees, and replay execution traces to understand how the harness arrived at specific actions.

Related

  • Dispatch
    Jeeves: Improving Decision Models with Reasoning and Diffusion Drafting

    Jeeves is a 9B reasoning classifier based on Qwen3.5-9B that outperforms Jev and Kev-9B on out-of-domain tasks by integrating a reasoning chain before decision-making.

  • Project
    vinilana/jev-gateway

    A local LLM gateway for coding agents that uses the fast Jev decision model to optimize tool selection, reducing cost and latency for the primary reasoning model.

  • Project
    PostHog/jeeves

    Jeeves is a reasoning-enhanced decision model based on Qwen3.5 that improves classification accuracy by generating a reasoning chain before outputting a final, calibrated decision.

  • Dispatch
    TypeSafe AI Jev System One Model: Fast, Structured Decision‑Making AI

    TypeSafe AI’s Jev, the first System One model, delivers frontier‑level intelligence for structured decision tasks at 40‑200× lower latency and 100‑500× lower cost than comparable LLMs, while guaranteeing type‑safe outputs and calibrated confidence scores.

  • Project
    BillionsBobby/JevRouter

    A fast agent routing layer that uses a specialized decision model to determine which tool or capability should be used next, reducing reasoning token waste and improving latency.