Harness Study
Personal infrastructure for building vendor-agnostic AI "agent harnesses" — versioned context layers (profile + skills + memory + workflows) that turn a generic AI into a consistent, specialized assistant.
Overview
Harness Study is a monorepo and ongoing research effort exploring "agent harnesses": vendor- and model-agnostic layers of structured context (profile + skills + memory + workflows) that turn a generic AI into a consistent, specialized assistant for a given domain. It hosts several harnesses (CS study, personal finance, and a harness-builder that scaffolds new ones via an interview flow) plus a library of reusable, shareable skills. This project reflects how I think about Applied AI — not just consuming models, but engineering the context that makes them reliable. I document what I learn in a public technical write-up.
Key highlights
- Designed vendor/model-agnostic "agent harnesses": versioned context layers (profile + skills + memory + workflows)
- Monorepo with multiple harnesses (CS study, finance) and a harness-builder that scaffolds new ones via interview
- A library of reusable, shareable skills across harnesses
- Reflects an engineering approach to Applied AI: building reliable context, not just prompts