Agents in the Wild

We track and analyze the activity and performance of autonomous code agents in the wild (on GitHub).

📊 Overview

We have been tracking all pull requests created since 2025-05-15. Our pull request database is updated hourly, capturing all newly created and closed pull requests since the last update. The table below provides an initial overview of the agents we track and their activity. Detailed breakdowns by specific PR attributes are presented in the following charts. More information on how we scrape, identify, and analyze this activity can be found at the bottom of the page or on our GitHub.

Agent	PRs Opened	PRs Closed	PRs Merged	Merge Rate
Human	7744447	6885121	5530970	80.33 %
Bot	4018325	3138791	2459886	78.37 %
OpenAI Codex	685430	634983	595401	93.77 %
Google Jules	83424	77818	72201	92.78 %
Github Copilot	31930	24786	18163	73.28 %
Cursor Agent	13249	11070	10060	90.88 %
Devin	8925	8231	6085	73.93 %

ℹ️ We calculate Merge Rate = Merged / Closed. Other trackers use Merged / Total, but we believe open PRs shouldn't count—they are either still under review or forgotten, neither of which indicates success or failure.

ℹ️ Different merge rates are heavily influenced by the level of human approval involved in the loop. For example, GitHub Copilot opens draft PRs without any prior approval, whereas OpenAI Codex only creates a PR after receiving human approval. Therefore, one should be very cautious when using merge rates as an indicator of agent quality.

📈 Daily Trends

Number of PRs

Number of PRs created on each day.

💡 Agents (primarily OpenAI Codex) contribute a significant volume, currently accounting for around 5–10% of public PRs. Merge rates vary significantly between agents, but this shouldn't be interpreted as a direct measure of agent quality, as the level of human input differs considerably.

⭐ Repo Popularity

We measure repo popularity by the number of stars the base repository of the PR has.

Portion of PRs

The percentage of PRs in each bucket.

💡 OpenAI Codex, which generates the most volume, seems to be primarily used on low-star (≤ 10 stars) repositories. This isn't generally true for all agents. Agent merge rates tend to drop significantly for more popular repositories, whereas human PRs tend to have more stable merge rates.

📐 Change Complexity

As a proxy for change complexity, we use the number of additions and deletions made in the PR. Each addition or deletion refers to one line.

Portion of PRs

The percentage of PRs in each bucket.

📂 Files Changed

Portion of PRs

The percentage of PRs in each bucket.

💡 Merge rates tend to decline slightly as more files are modified. Agents are not afraid of editing multiple files.

± Additions/Deletions

Here we analyze the ratio between additions and deletions. We define AD = additions / (additions + deletions). A high AD suggests a lot of new code, while an AD around 0.5 indicates refactoring. An AD close to zero means a lot of code was deleted and very little was added.

Portion of PRs

The percentage of PRs in each bucket.

💡 Agents tend to add more code than they delete (potentially adding features), while humans are more likely to refactor. Merge rates remain stable across cleanup, refactoring, and additions.

🔣 Repository Language

Here we analyze the primary language of the base repository. GitHub assigns each repository one primary language that makes up most of the code in the repository.

Portion of PRs

The percentage of PRs in each bucket.

🔣 Change Language

Here we analyze the primary language of the change. Since GitHub does not provide this, we determine the primary language by the file extensions of changed files. We associate each file extension with a language and determine the most changed language. Certain filenames (e.g., package-lock.json, *.lock) are excluded.

Portion of PRs

The percentage of PRs in each bucket.

📦 Data

We track all opened and closed GitHub pull requests through the GitHub API, and analyze each pull request to determine if it was authored by an autonomous code agent.

To classify each pull request, we use the following rules:

Human: All PRs that do not match any known agent pattern and are authored by a User account.
Bot: All PRs that do not match any known agent pattern and are authored by a Bot account.
OpenAI Codex: PRs where the head branch starts with codex/.
Google Jules: PRs where the first commit is authored by google-labs-jules[bot].
GitHub Copilot: PRs where the head branch starts with copilot/.
Devin: PRs authored by user devin-ai-integration[bot].
Cursor Agent: PRs where the head branch starts with cursor/.

✨ Ideas?

Whether you have ideas for new autonomous agents, additional insights, or anything else, please create an issue on our GitHub repository or email insights@logicstar.ai.

🌎 Open Source

The source code and documentation explaining how we scrape data, identify agents, and more can be found on our GitHub repository. Like the project? Give it a star!