Building an AI Agent Team That Reviews Like Real Koreans (Harness Fork + NVIDIA's 1 Million Personas)

Existing AI agent teams, even when composed of multiple agents, have a limitation in that their tone and perspective are similar, lacking real diversity in feedback. To solve this, a harness fork project has been released that clothes agents in real Korean workplace personas (job role, generation, region, family situation, etc.) using the NVIDIA Nemotron-Personas-Korea dataset (1 million rows).

While maintaining the structure of the existing revfactory/harness, this project adds the following three core skills:

• korean-persona-search: Dynamically searches and samples diverse personas based on job role, region, age, education, and more.
• korean-voice-adapter: Adjusts speech style (formal/polite endings, etc.) by reflecting Korean workplace culture (reporting manners, meeting styles) and a vocabulary dictionary covering 13 industries.
• korean-persona-harness: Acts as a meta-orchestrator managing a 5-person sub-agent pipeline that spans from scenario analyst to diversity QA.

Validation results show that, compared to a general agent team, voice distinctiveness is very high, and Korean-specific workplace manners (mentoring, indirect expressions) and personal details (mentions of family schedules, etc.) are naturally reflected, enabling much more three-dimensional reviews.

This technique offers high value across various fields such as code review, virtual user interview simulation, marketing copy review, and UX research, and is compatible with Claude Code and Codex CLI.