Editorial directory · v0.1

The directory of software delegates.

AI systems that accept an assigned software task, work inside a bounded development environment, and return a reviewable pull request. The table below compares every product that meets the bar today.

Entries: 6
Vendors: 6
Reviewed: 2026-05-12

Jump to entries Read the paper

Compare

Click a row for the notes below. Hover a column header for the definition.

Tool	Input	Models	Iteration	Output	Access	Source
Vroni by Vroni	Idea or full spec	Any provider (OpenAI, Anthropic, Google, others)	Iterates until quality gates pass	Pull request	Early access	Docs
Claude Code GitHub Actions by Anthropic	Detailed prompt or issue	Anthropic only (API, Bedrock, Vertex)	Single agent loop	Pull request	Public · API key required	Docs
Cursor Cloud Agents by Cursor	Task description	Multi-provider (Claude, GPT, others)	Single agent loop	Pull request	Cursor Pro, Business or Enterprise	Docs
GitHub Copilot coding agent by GitHub	Issue or Copilot chat prompt	GitHub-hosted models	Single agent loop	Pull request	Copilot Pro+, Business or Enterprise	Docs
Google Jules by Google	GitHub issue description	Google only (Gemini)	Single agent loop	Plan and diff	Public	Docs
OpenAI Codex web by OpenAI	Task prompt or @codex mention	OpenAI only	Single agent loop	Pull request	ChatGPT Plus, Pro, Business, Edu or Enterprise	Docs

Notes

One paragraph per tool, same order as above.

Vroni

by Vroni

vroni.com

Vroni wraps coding agents inside a larger engineering workflow rather than exposing a coding agent directly. A task is assigned through a GitHub issue or the Vroni web UI; the system plans the change, picks tools, runs build and test loops, repairs failures, deploys preview environments, exercises the running application, and switches between models from different providers when a step stalls. It stops when its own quality checks pass.

Input: Idea or full spec
Iteration: Iterates until quality gates pass
Models: Any provider (OpenAI, Anthropic, Google, others)
Output: Pull request
Assignment: GitHub issue, web UI
Runs in: Dedicated cloud VM
Authority: Full engineering workflow
Access & pricing: Early access · Per active runtime minute

Claude Code GitHub Actions

by Anthropic

code.claude.com

Anthropic packages Claude Code as a GitHub Action that runs on standard GitHub-hosted runners. A mention of @claude on an issue or pull request, or an issue assigned to the action, triggers a run. The action reads the repository, performs the requested change, and opens a pull request. Configuration sets the model (Claude Sonnet by default, Opus available by parameter), the maximum number of agent turns, and the allowed tools.

Input: Detailed prompt or issue
Iteration: Single agent loop
Models: Anthropic only (API, Bedrock, Vertex)
Output: Pull request
Assignment: @claude mention, issue assignment
Runs in: GitHub-hosted runner
Authority: Single agent loop
Access & pricing: Public · API key required · Anthropic API usage + GitHub Actions minutes

Cursor Cloud Agents

by Cursor

cursor.com

Cursor's background agents extend the Cursor editor into a remote environment. Work can be assigned from the editor, the Cursor web app, Slack, Linear, or GitHub. Agents run in isolated machines and report status; users can send follow-ups or take over a session. The output is a pull request when the work targets a repository.

Input: Task description
Iteration: Single agent loop
Models: Multi-provider (Claude, GPT, others)
Output: Pull request
Assignment: Editor, web, Slack, Linear, GitHub
Runs in: Isolated Cursor VM
Authority: Single agent loop
Access & pricing: Cursor Pro, Business or Enterprise · Included in Cursor plans

GitHub Copilot coding agent

by GitHub

docs.github.com

GitHub's coding agent works inside the GitHub ecosystem itself. A user assigns an issue or sends a request from chat; the agent researches the repository, plans changes, edits files, and opens a pull request that integrates with GitHub's review and CI flow. Execution happens on GitHub-managed infrastructure.

Input: Issue or Copilot chat prompt
Iteration: Single agent loop
Models: GitHub-hosted models
Output: Pull request
Assignment: GitHub issue, Copilot chat
Runs in: GitHub-managed infra
Authority: Single agent loop
Access & pricing: Copilot Pro+, Business or Enterprise · Included in Copilot plans

Google Jules

by Google

jules.google

Jules is Google's asynchronous coding agent. It connects to GitHub, takes work in the form of an issue, and runs in a secure cloud VM. The artifact is a plan plus a diff, returned for review and merge inside the user's existing workflow.

Input: GitHub issue description
Iteration: Single agent loop
Models: Google only (Gemini)
Output: Plan and diff
Assignment: GitHub issue
Runs in: Google cloud VM
Authority: Single agent loop
Access & pricing: Public · Free tier + paid plans

OpenAI Codex web

by OpenAI

developers.openai.com

OpenAI's web-hosted Codex agent accepts work either through its own web interface or through @codex mentions on GitHub issues and pull requests. Work runs inside OpenAI's cloud environment with read/write access to the connected repository and proposes the result as a pull request. The execution is one Codex session per task; the artifact is reviewed in GitHub like any other PR.

Input: Task prompt or @codex mention
Iteration: Single agent loop
Models: OpenAI only
Output: Pull request
Assignment: Web UI, @codex mention
Runs in: OpenAI cloud sandbox
Authority: Single agent loop
Access & pricing: ChatGPT Plus, Pro, Business, Edu or Enterprise · Included in ChatGPT plans

Methodology

What qualifies, where the data comes from, who's behind this list.

Criteria

Every entry meets these four conditions.

01 Task entry. A structured way to assign software work — issue assignment, chat prompt, PR comment, IDE action, or similar.
02 Bounded environment. Scoped repository, runtime, and tool access for the assigned task.
03 Implementation authority. Carries out at least one development action without re-prompting — edit files, run commands, push a branch.
04 Reviewable artifact. Diff, branch, pull request, log, or test report, with a human acceptance step before merge.

Autocomplete, prompt-only coding chat, and fixed CI pipelines do not meet the bar on their own.

Sources

Public vendor documentation is the primary source for every entry. Each row shows the date of the most recent documentation review. When a vendor changes scope, authority, or output in a way that affects category fit, the row is updated.

Entries are not commercial placements. There is no fee for inclusion and no paid ordering.

Disclosure

Maintained by Vincent Schmalbach, founder of Vroni, which appears in the directory.

The category was defined in writing before Vroni was placed against the criteria; the definition is anchored in the working paper Delegated Software Agents. The four-property test above applies to every entry the same way. Corrections via the maintainer's site.