⬢github Python · 4.9K ★ +4.1K since we first saw it · pushed 4 d ago · Apache-2.0
google-research/harness2
Harness² is a Python framework from Google Research for recursively improving the 'harness' around an LLM agent—prompts, tools, memory, skills, guardrails, and more—while keeping model weights fixed. It refines the harness for each task via a Propose–Probe–Compose loop and accumulates lessons across tasks, reporting gains on benchmarks like JobBench, LAB, and WorkBuddy-Bench.
Why now: It appeared among the most-starred new GitHub repositories shortly after release, backed by an arXiv paper, demo video, and reported benchmark improvements.
Who it is for: AI researchers and engineers building or tuning LLM agent systems who want to optimize harnesses without fine-tuning models.
Stars over our 37 snapshots: 884 to 4.9K, since 1 d ago.
Where people talked about it
- ⬢github new repos, most starred 6 min ago
- ⬢github new repos, most starred 18 h ago
API: https://socialmediatrends-api.osmike.com/v1/repos/google-research/harness2