A report card for the guardrails around your AI coding agent

What it does harness-score is a CLI tool and GitHub Action that scans a repository for the “harness” around AI coding agents: the AGENTS.md files, .cursor/rules/, test runners, CI pipelines, and guardrails. It assigns a maturity level from L0 (unharnessed) to L4 (self-correcting) based on 36 deterministic filesystem checks across six dimensions, producing the same result every run with zero network access or LLM calls.

The interesting bit Instead of judging whether your tests are meaningful or your rules are fresh, it treats maturity as infrastructure: does the file exist, does the CI job run, is the pre-commit hook installed? That makes the score a binary, gatable signal you can enforce in a pipeline—something a linter wishes it could claim about behavior.

Key highlights

• 108-point breakdown across six dimensions: Context, Skills, Hooks, Sensors, CI, and Hygiene
• Maturity levels gate on shape, not just points—beautiful docs with zero tests still lands you at L1
• Fully deterministic: same commit, same score, no network requests, no LLM calls
• Built-in badge generation and a --min-level flag for CI gating
• Supports Cursor, Claude Code, Windsurf, Cline, Continue, and others via OR semantics

Caveats

• The scanner only verifies that artifacts exist and parse; it cannot tell if your rules are stale or your tests actually catch bugs
• Editor plugins beyond Cursor are still in progress
• Functional correctness and team review culture live outside the tree, so a high score is necessary but not sufficient

Verdict Worth running if you’re “vibe-coding” with AI agents and want a ranked todo list for hardening your repo. Skip it if you already have rigorous agent governance and don’t need a numbered badge to prove it.