Test suites: why AI coding agents need tests they can find

What counts as a test suite?

• Full pass: a tests/, test/, __tests__/, spec/, specs/ or src/test/ directory.
• 70%: no such directory, but test-named files turn up within three folder levels, such as *.test.ts, test_*.py, *_test.go or *Test.java.
• Fail: neither. We check that tests exist, not that they pass.

How checks become a score: methodology.

How do I make tests easy for AI agents to find?

Keep tests in a conventional directory and put the command that runs them in AGENTS.md. Tests next to the source score 70%, because an agent has to search for them.

Which AI coding agents weigh tests most?

• Aider14% of its score
• Claude Code13% of its score
• Pi12% of its score
• Kimi Code CLI12% of its score
• Devin11% of its score
• Codex10% of its score
• Gemini CLI10% of its score
• OpenHands9% of its score
• Cursor8% of its score

Pass rate by language

• Haskell7 of 7 (100%)
• Ruby9 of 10 (90%)
• Swift9 of 10 (90%)
• JavaScript13 of 15 (87%)
• Elixir6 of 7 (86%)
• C8 of 10 (80%)
• Clojure4 of 5 (80%)
• PHP9 of 12 (75%)
• Zig4 of 6 (67%)
• Python37 of 57 (65%)
• C++13 of 21 (62%)
• C#4 of 8 (50%)
• Rust22 of 45 (49%)
• Dart3 of 7 (43%)
• Go16 of 40 (40%)
• Java10 of 28 (36%)
• Scala2 of 6 (33%)
• TypeScript24 of 82 (29%)
• Kotlin1 of 12 (8%)