JevHarness
An LLM writes a task-specific Jev harness, improved with rewards and traces; a Pokemon harness rose from 25% to 75% wins.
Overview
JevHarness splits work between a strong authoring LLM and Jev. The LLM writes a task-specific harness: code that turns observations into useful features, constructs Jev questions and criteria, and combines structured answers into actions. Once the harness is frozen, execution runs that code and its Jev calls without the authoring LLM on every decision, so the LLM's reasoning strategy is preserved in explicit code, state, instructions and control flow while Jev makes fast, fuzzy, context-sensitive decisions from computed facts.
Optionally, the harness is improved using rewards and complete execution traces through reflection and GEPA evolution, with the authoring LLM revising code, questions, graph and memory during development. In the README's Pokemon battle example, five reflection rounds improved the selected harness's evaluation win rate from 25% (3 of 12) to 75% (9 of 12), with the round-3 candidate retained as best; the author notes the evaluation set was used for selection. An interactive demo compares initial and evolved harnesses, a latency section explains the speed trade-off, and it installs as a Claude Code skill. Documentation is in English and Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Task-specific Choice, Score and Noul questions authored into each harness (e.g. Pokemon battle moves) | — |
How it works
- Authoring LLM writes a harness of features, Jev questions and action logic
- Run the harness; Jev makes per-step decisions
- Optionally reflect on rewards and traces to evolve the harness
- Freeze the best harness for fast execution
Strengths
- Deep reasoning at development time, fast decisions at runtime
- README reports win rate rising from 25% to 75% in the Pokemon example
- Installs as a Claude Code skill
Caveats
- Evaluation set also used for selection in the reported result
- Small evaluation (12 scenarios)
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Claude Code skill / harness runtime
- Has code
- yes
Use cases and integrations
How curated lists describe it
LLM-authored task-specific Jev harnesses with optional full-trajectory reward…
wh000wh000/awesome-jev-liveLLM-authored task-specific Jev harnesses with optional full-trajectory reward reflection and GEPA evolution.
hellogumbo/awesome-jev
Frequently asked questions
Does the LLM run on every decision?
No, once the harness is frozen only its code and Jev calls run.
How much did evolution help?
In the Pokemon example, evaluation win rate went from 3/12 to 9/12 after five reflection rounds.
How do I use it?
Install it as a Claude Code skill and let the agent author a harness for your task.
Related projects
Multi-agent orchestration layer for Claude Code with opt-in Jev hooks for fast judgment points, decision caching and egress controls.
PluginproductionTypeScriptAgent plugin that asks Jev whether the session is at a safe boundary to /compact and can trigger it on Pi and Claude Code.
PluginbetaTypeScriptCoding-agent hook where Jev decides whether a mid-turn message should steer the running task, be queued, or interrupt it.
PluginbetaPythonOne skill and MCP server for Claude Code, Codex and Hermes where Jev gates large-model turns, routes models and guards Git steps.
MCP serverbetaPythonRuns Claude Code on Opus 5.5 with Jev re-picking reasoning effort after every tool batch, via per-message effort so the cache survives.
ProxybetaTypeScriptCLI where several coding agents draft implementation plans and Jev scores them, decides on cross-review and picks the finaliser.
ProjectbetaTypeScript