References
The sources this guide consolidates, roughly in order of influence.
Harness engineering
- Harness engineering for coding agent users — martinfowler.com, April 2026. The article that framed the discipline for teams using agents: guides vs. sensors, computational vs. inferential checks, the three regulation dimensions (maintainability, architecture fitness, behavior), "keep quality left", and harnessability as a property of codebases.
- Harness Engineering — first thoughts — the earlier memo in the Exploring Gen AI series where the term takes shape.
- Improving Deep Agents with harness engineering — LangChain, 2026. The empirical case: 52.8% → 66.5% on Terminal Bench 2.0 without touching the model. Self-verification loops, context assembly middleware, loop detection, and the reasoning sandwich.
- deepagents — LangChain's open-source "batteries-included agent harness"; useful to read as a concrete harness implementation.
Cursor documentation
- Rules —
.cursor/rules/*.mdc, frontmatter, AGENTS.md. - Agent Skills — the SKILL.md standard and how Cursor loads skills.
- Hooks — hooks.json, the full event list, permission decisions.
- Plugins — plugin structure, manifest, installation modes.
- Cursor Marketplace and the plugin spec repository — how plugins are packaged, reviewed, and distributed.
Adjacent work
- Anthropic's writing on effective agents and agent SDKs — much of the context-engineering advice transfers directly to Cursor harnesses.
- The Agent Skills open standard adopted across tools — the reason skills written for Cursor are portable.
- ArchUnit, dependency-cruiser, import-linter — architecture fitness functions referenced in chapter 4.
About this project
harness-score (the scanner, this guide, the Cursor plugin, and the GitHub Action) is open source under MIT: github.com/paladini/harness-score. The repository dogfoods everything it teaches — it scores L4 on its own scanner, and its CI gates on --min-level 4.