What you can compare with today
Open today: public repositories. The scores of 204 public open-source repositories, scanned on 2026-11-14.
Open today: a new installation. What the scan grades for each coding agent before a team adds a single file.
Opens later: results that other teams share. A figure appears when a group holds at least 11 shared results. Until then you see the count.
See the comparison for your result
python3 assess.py --benchmark-file results/results.json
AEI Scan prints your scores beside the 3 comparisons below. The request for the figures holds no part of your result.
Public repositories
On 2026-11-14 we scanned 204 public open-source repositories that commit coding-agent files. They gave 286 results, because a repository with files of 2 coding agents gives 2.
| Coding agent | Results | Median overall score | Middle half of the results |
|---|---|---|---|
| Claude Code | 141 | 19.1 | 16.6 to 21.6 |
| VS Code with Copilot | 73 | 36.7 | 28.3 to 36.7 |
| Antigravity | 40 | 50.0 | 25.0 to 56.3 |
| Cursor | 27 | 36.7 | 31.1 to 43.4 |
| Codex CLI | 5 | no figure: fewer than 11 results |
The scan read each repository only and no answers were given, so these scores come from the 35 scan checks alone. For a like-for-like score of your own, run AEI Scan with --repository-only and no answers. A score for one coding agent is not comparable with a score for another, because the scan reads a different number of checks for each.
A new installation
We ran the scan for each coding agent on an empty repository and an empty home folder. This is what the scan grades before a team adds a single file.
| Coding agent | Scan checks | Passed | Failed | Not read |
|---|---|---|---|---|
| Claude Code | 35 | 0 | 30 | 5 |
| Codex CLI | 35 | 0 | 18 | 17 |
| Cursor | 35 | 0 | 19 | 16 |
| Antigravity | 35 | 0 | 8 | 27 |
| Grok Build | 35 | 0 | 1 | 34 |
| VS Code with Copilot | 35 | 0 | 17 | 18 |
| Xcode | 35 | 0 | 1 | 34 |
No coding agent passes a scan check before a team configures it. Every point of a score comes from work the team did.
Opens later: results that other teams share
The results site keeps the results that people upload. A member shares a result in the benchmark by ticking a box when claiming it. The box is empty until the member ticks it, and sharing can be turned off again.
You see the median and the middle half, overall and for each of the 16 dimensions. Each member counts once per coding agent, with the latest result. A group needs at least 11 results before any figure appears; below that you see the count. The sample results from Prosights and the result of Kyberna are left out.
Read the shared figures. Until 11 members share a result, that page shows a count and no figure.
Members choose to share. The shared results are therefore not a sample of all teams, and their median is not the median of the industry.
For investors: several companies side by side
A member of the results site can keep a portfolio: the companies whose results that member reads side by side.
You make an invitation link for each company and send it. A person at the company signs in, reads what will be shared and with whom, and agrees or declines. Nothing is shared before that.
You then read each company's latest scores, level, coverage and claim gap, and its earlier runs.
For the portfolio you read the lowest, the median and the highest score in each dimension, and the dimensions where most companies score lowest.
You do not read the checks, the reasons, the file names or a company's counts for each model. A company can share those as a second, separate choice.
A company can withdraw at any time. The portfolio then no longer shows its results.
You can save the table as a CSV file and print a summary. You cannot add a result for a company.
A member can also say who owns the organization: a venture capital fund, a private equity fund, public shareholders or other private owners. A venture-capital-backed member can add the stage of its latest financing round. The benchmark shows a median for such a group once the group has 11 results.
Each company runs the scan itself and gives its own answers. The site does not verify the answers. A score is not a statement about a company's value.
For investors: before a deal
A member of the results site can ask a company for its results before investing in it or buying it. The site calls this a pre-deal assessment.
You make a request under a code name, say what you ask for and set a last day, then send the request link. The company runs AEI Scan on its own machines, for as many repositories as it names, and uploads the results. Uploading needs no account.
You read nothing until a person at the company signs in and agrees. That person first states which repositories are in the results and which are left out. The report prints this as the company's statement.
Every graded check in the report has 1 of 4 evidence grades: Declared, Scanned, Witnessed, or Verified by assessor. The report gives the share of the score at each grade.
An assessor whom Prosights has approved, and whom you name, can watch a run and grade the checks that need an artifact shown.
The report starts with a page for an investment committee, then gives the dimensions in a table, the findings by severity with the rule in full, the questions for management and a remediation plan. The detail follows as an appendix, and the report ends with what it does not show.
The company sees who reads the report and every time it is opened. The company can withdraw at any time. On the last day the site deletes the results.
Read a sample pre-deal report, or the same report as a PDF. The sample comes from 3 made-up repositories.
A run that nobody outside the company watched can be staged. The report makes no claim about value, returns or price.
What a comparison does not show
A higher score than the median does not show faster delivery or better code. No study read for the index shows that.
The public repositories are open-source projects. They do not describe private company repositories, and settings on a maintainer's own machine are not in them.
How the public sample was drawn
The list of repositories is the published data of a study: Galster and others, "Configuring Agentic AI Coding Tools: An Exploratory Study" (arXiv 2602.14690; data). The list holds 2,853 repositories on GitHub with files of the coding agents the study searched for. Other repositories are not in it.
We drew 300 of them at random and scanned 204.
Not scanned: 36 larger than 300 MB, 9 that could not be copied within 90 seconds, and 51 with no coding-agent file that the scan knows.
No repository is named in the figures, and no figure is given for fewer than 11 results.
Compare a result of your own
Write to info@prosights.com for AEI Scan, run it in a repository of your own, then run the comparison command above.
Prosights.ai