AI Explainedby Milad Fahmy

AI Explained

Background coding agents

Assign a ticket before lunch, and a pull request is waiting when you're back. How background coding agents work in cloud sandboxes, and how to use them well.

Milad FahmyNov 7, 20262 min read

On this page
  1. Local vs background
  2. How it works
  3. Good tasks vs bad tasks
  4. Stay in control
  5. A good rhythm

You assign a ticket before lunch. When you're back, a pull request is waiting. That's a background coding agent: it works on its own, in the cloud, while you do something else.

Local vs background

A local agent works in your terminal or editor, and you watch each step. A background agent takes the task, works independently, and reports back when it's done.

GitHub Copilot's cloud agent, OpenAI Codex, Claude Code on the web and Google's Jules all work this way.

How it works

  1. A fresh cloud machine per task. OpenAI describes each Codex task running "in its own cloud sandbox environment, preloaded with your repository", with dependencies installed.
  2. Isolation. Because every task has its own machine, you can run several at once without them colliding.
  3. Limited network. Internet access is restricted by default (GitHub uses a firewall, for example), so the agent can't reach just any site.
  4. A plan first, sometimes. Jules, for example, shows a plan you approve before any code changes.
  5. Work without supervision. The agent edits files, runs the tests and fixes what fails.
  6. A pull request at the end. The result comes back as a branch or PR, so you review it like a teammate's work.

Many of these agents read an AGENTS.md file in your repo to learn how to build, test and follow your rules.

Good tasks vs bad tasks

Good: small and clear. A bug with steps to reproduce, a test to add, a dependency bump, a small refactor.

Bad: vague requests come back wrong, and big redesigns still need you in the loop.

Stay in control

  • Review every pull request. The agent did the typing; you own the merge.
  • Protect your main branch with rules, so no agent can push there directly.
  • Keep secrets out of the agent's environment unless the task needs them.

A good rhythm

Hand off the small tasks in the morning, and review the pull requests after lunch.

Delegate the task. Keep the review.

Sources

#background agents#coding agents#codex#github copilot#claude code

Watch