HKUDS

HELIX

Name: HELIX
Author: HKUDS

Data community

SOURCE https://github.com/HKUDS/HELIX

Description

Helix's mission: optimize harnesses for models, then automatically turn harness-generated rollouts into trainable agent data.

Installation

This entry records only its repository, not the path inside it, so there is no exact command to give. Open the source below and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

README

English | [中文](README_zh.md)

Helix: Let Models and Harnesses Co-Evolve

Helix's mission: optimize harnesses for models, then automatically turn harness-generated rollouts into trainable agent data.

![TypeScript](https://img.shields.io/badge/TypeScript-5.9-blue) ![Node](https://img.shields.io/badge/Node-20%2B-339933) ![Benchmarks](https://img.shields.io/badge/Benchmarks-LiveCodeBench%20%7C%20SWE--bench-orange) ![Status](https://img.shields.io/badge/Status-research%20prototype-purple)

What Is Helix?

Helix is a workspace for general-agent harness composition, evaluation, and data-loop construction. It makes it easy to build and validate different harnesses, while turning existing harnesses into reusable modules.

What is a harness?

shell + session/hooks + config + prompt + tools + turn loop + acceptance + runtime policy

The complete harness shapes how an agent plans, calls tools, manages context, accesses services, recovers from failure, and produces verifiable work. Many existing harnesses already capture strong engineering judgment, but their combinations are often similar. Helix asks whether we can break them into lego blocks, recombine them, and search for newer, stronger harnesses.

Helix asks:

For a model, how do we choose or compose the harness that fits it best?

For a scenario, how do we automatically build a data loop that pushes the
model's capabilities as far as possible while collecting multiple kinds of
rollout data for the next round of model iteration?

News

2026-07-19: Added the full search setting, making harness search stronger.
2026-07-17: Helix outperformed all upstream harnesses on the LiveCodeBench subset.
2026-07-16: Helix outperformed