Back

Field study

119 open-source repos, scored for coding agents

crag, my open-source tool, checks whether a repo tells a coding agent how to work in it. I ran it on 119 widely used projects. 49 of them had nothing for an agent to read.

Repos
119
Tool
crag 0.6.1
Data
Public leaderboard, Jul 2026
Published

Result

What I found

A coding agent starts with no idea of a repo's test command, its commit rules or which folders it shouldn't touch. Some projects write that down in a file at the root, like AGENTS.md or CLAUDE.md. Many don't.

Of the 119 repos, 49 had no such file, which is 41 percent. All 49 sat in the bottom two grades, and no repo in those grades had one.

Every number on this page comes from one fetch of the public leaderboard API in July 2026.

Method

What crag checks

crag looks for the instruction file. It counts the quality gates the repo defines, meaning the commands an agent should run before it commits. And it checks that the written guidance matches what the config and CI actually do.

Each result goes on the leaderboard with a grade and the counts behind it.

I chose the repos myself, picking projects that are widely used, so the sample leans toward well-maintained code.

The score only covers this scaffolding. It says nothing about whether the code is good, so an excellent library can still score low.

Numbers

The grades

Recomputed from the raw leaderboard payload, July 2026.

119repos scored
35 / 35in grade A and grade B
47in grade C, the largest group
41%with no instruction file for agents

Distribution

Grade C is the largest group

47 repos in grade C, 35 each in grades A and B, and 2 in grade D. The 47 in grade C and the 2 in grade D are the 49 repos with no instruction file.

Details

What the top repos have in common

Thirty repos scored a perfect 100. Every one had at least one instruction file, and most had two or three.

The median repo defines 38 quality gates. Only 10 of the 119 had any drift between what the instructions say and what the config does.

Sources

  1. 119 repos on the leaderboard; grade distribution A 35, B 35, C 47, D 2; 30 perfect scores; 49 repos (41 percent) with zero agent-instruction files; median 38 quality gates; 10 repos with any drift; crag version 0.6.1. All recomputed from the live payload.crag leaderboard API (public JSON)accessed 2026-07-06
  2. The public leaderboard page listing every audited repo, its grade, and the raw countscrag.shaccessed 2026-07-06
  3. crag, the governance and agent-readiness scoring tool, and its documentationcrag.shaccessed 2026-07-06
  4. crag documentation: how the audits and scoring workcrag.shaccessed 2026-07-06