Harness Framework

A declarative, skill-based harness for rapidly composing LLM agents.

A declarative agent harness that turns hard-coded multi-agent pipelines into composable, configuration-driven skills.

  • LLM semantic routing selects the right skill from natural-language intent, so new capabilities are added by writing a skill spec rather than new code.
  • Progressive disclosure loads only the context a task needs, keeping prompts small and routing cheap.
  • A unified tool registry, permission interception, and a sandboxed execution layer make skills safe to run in production.
  • A conversational skill studio lets non-engineers author and validate new skills end-to-end.

Built during my AI research internship at TikTok (ByteDance) for large-scale e-commerce governance.