GitHub radar
New abilities for Claude Code using LongHorizon-Harness
A tool built by the AMAP-ML team for providing an execution & verification layer on top of Claude Code and Codex, boosting task success rate from ~50% to ~80% on benchmarks. Climbed to the #1 spot on Hugging Face Daily Papers’s weekly ranking on August 6th, 2026.
LongHorizon-Harness is an open source state management and execution framework developed by AMAP-ML team to run atop of other tools such as Claude Code CLI / Codex CLI. It supports a Manager, an Executor and an Auditor role to decouple your knowledge about what you want to achieve (goal) and your verified history, and execute each step with a clean context while verifying its result before adding it to your verified history. LongHorizon-Harness improves WeaveBench’s task success rate from ~50% to ~80%. It also triples OSWorld 2.0’s desktop full task completion rate. LongHorizon-Harness reached the #1 spot on Hugging Face’s Daily Papers weekly ranking on August 6th, 2026.
Why a vibe-coder should care
LongHorizon-Harness gives you an organizational structure to manage your verified history when building increasingly complex things using the same model such as coding, testing, debugging, deploying, etc…
▌ More finds