Someone on your team has started saying "just give it to the agent". An issue goes to Copilot or a task goes into Codex, and later a pull request appears that nobody typed. If you're deciding whether that's safe, or worth paying for, it helps to know what happened in between.
The four agents compared here all expect a developer to read what comes back. If the person asking is a product owner who'd rather click through a change than read a diff, where Refinar fits at the end covers that case.
What a background coding agent does
A background coding agent is an autonomous AI coding agent that works on its own copy of your repository in the cloud while you do something else, then hands back a branch or a pull request. The big four follow roughly the same sequence.
- It starts a fresh container or virtual machine and clones your repository at the branch you chose. Your real repository only ever receives a branch.
- A setup script installs dependencies, and the agent reads any house rules you've left, such as
AGENTS.md(which file each agent reads). - It loops on its own. Edit, run a command, read the output, edit again, until the task looks done or it gives up.
- It runs the tests and linters it can inside the copy. Nobody has necessarily watched the app run.
- It pushes a branch, and the work reaches a person as a pull request.
The four differ mostly in where the copy runs and what the agent may do to GitHub.
| Agent | Runs in | Hands back | Cheapest plan with it |
|---|---|---|---|
| Codex cloud | A container OpenAI runs | A diff, then a PR you open | ChatGPT Plus, $20/month |
| Claude Code on the web | An isolated VM Anthropic manages | A pushed branch, then a PR or draft | Claude Pro, $20/month |
| Cursor Cloud Agents | An isolated VM with a dev environment | A branch and PR, with screenshots | Paid plan, plus API-priced usage |
| Copilot cloud agent | A GitHub Actions environment | A copilot/ branch and a draft PR | Copilot Pro, $10/month |
Prices from each vendor's site, 29 September 2026.
OpenAI Codex (cloud)
Codex cloud lives at chatgpt.com/codex. You connect GitHub (GitLab is in beta) and create an environment per repository. When a task starts, Codex creates a container and checks out your branch. Setup scripts get internet access; the agent itself doesn't unless you switch it on, and secrets are removed before the agent phase begins. That's a sensible default for something that reads issues written by strangers. You get a summary and a diff at the end, and you decide whether to open a pull request.
Codex cloud comes with ChatGPT Plus, Pro (from $100), Business and Enterprise. An API key alone doesn't include it, and cloud tasks spend the same allowance as local ones.
Claude Code on the web
Claude Code is Anthropic's agentic coding tool, and its web version at claude.ai/code runs each session in an isolated, Anthropic-managed VM that keeps going after you close the laptop. Access comes through the Claude GitHub App, and in Anthropic-hosted environments your GitHub credentials never enter the VM. At a stopping point Claude pushes a branch, and from the diff view you open a full or draft PR. Auto-fix can then watch that PR and push fixes for failing CI or review comments.
Claude's Free plan doesn't include it. Pro ($20 a month, $17 billed yearly), Max (from $100), Team and eligible Enterprise seats do. The VM carries no separate charge, but sessions share rate limits with the rest of your Claude usage.
Cursor background agents, now Cloud Agents
Cursor renamed background agents to Cloud Agents. Each runs in an isolated VM with a full development environment, cloned from GitHub, GitLab, Azure DevOps or Bitbucket Cloud, and works on a separate branch it pushes for handoff. It tries hardest of the four to show its work. Agents attach screenshots and videos showing how they checked a change, and you can take over the agent's remote desktop to try the software yourself.
You need a paid plan (Individual is $20 a month, Teams $40 per user), and cloud agents are billed on top at API pricing for the model you pick, with a spend limit set on first use.
GitHub Copilot coding agent, now Copilot cloud agent
GitHub renamed Copilot coding agent to Copilot cloud agent in April 2026. Assign an issue to Copilot or mention @copilot on a pull request, and it works in an ephemeral environment powered by GitHub Actions. GitHub documents the strictest rules of the four. Only people with write access can trigger it. It pushes only to its own copilot/ branch and opens drafts that a human has to mark ready and merge, and whoever asked for the change can't approve it. By default, your Actions workflows don't run on its PR until someone with write access clicks Approve and run workflows. Sessions stop at 59 minutes, and it works only on GitHub.
Every paid Copilot plan includes it, starting with Pro at $10 a month. It uses Actions minutes and AI credits, and Business or Enterprise admins must switch it on.
Which is the best AI coding agent for you
- The team already pays for ChatGPT: Codex.
- Your developers already use Claude Code: Claude Code on the web.
- You want evidence of the change beyond a diff: Cursor, for its screenshots and remote desktop.
- Work starts as GitHub issues behind branch protection: Copilot.
- The person asking doesn't code: see the best vibe coding tools instead, or Lovable vs Claude Code vs Codex if you're torn between an app builder and an agent.
Whichever you pick, the agent fills every gap in the ticket with a guess. The PRD generator helps close those gaps first, and the pull request explainer reads a public PR back to you in plain English.
Where Refinar fits
The four agents above leave the reviewing to a developer. Refinar is a background agent for the person on the other side of that review. A product owner describes a change in plain words, and the entire app starts up in a sandbox, connected to the APIs and services it uses in production, so they judge the result by clicking through it. Writes to those services are flagged for approval. What reaches GitHub is a draft pull request, and the sensitive categories (logins, payments, customer records) can't go further without a developer (how it works).
If a preview you can click suits your team better, try Refinar on one repository and compare its draft pull request with what your current agent hands back.