Varun Vemulapalli

Day 76 of 100 · Phase 9: Claude for builders

Recap of Days 55–75: comparing the coding agents

2 min readBy Varun Vemulapalli

Three weeks, three coding agents: GitHub Copilot, OpenAI Codex, Claude Code.

Honest summary: they now do much the same core job. The real difference is where each one lives in your day.

Video explainer · 1:39
Key takeaway

Copilot, Codex and Claude Code share the same core skills; pick by where your work happens and keep a human reviewing every change.

Day 76 visual: Recap of Days 55–75: comparing the coding agents
The visual from the original post.

What we covered (Days 55–75)

GitHub Copilot (55–60): completions and Chat in your editor, agent mode across files, custom instructions and MCP servers. The agent you assign an issue to was called the coding agent on Day 58; GitHub's docs now call it Copilot cloud agent.

OpenAI Codex (61–65): one agent across the CLI, IDE extension, Codex Cloud and the ChatGPT apps, guided by AGENTS.md, used for tests, refactors and review.

Claude (66–75): claude.ai Projects and artifacts, the API and tool use, Claude Code (CLAUDE.md, skills, MCP, subagents, hooks, permissions) and the Agent SDK.

What all three share

  • Read a repo, edit across files, run commands and tests.
  • Take standing instructions from a file in the repo.
  • Connect to other tools through MCP.
  • Need a human to review every change.

Where each one lives

  • GitHub Copilot: in your editor and on GitHub itself. A natural fit when work runs through GitHub issues and pull requests. Plans range from Copilot Free to Business and Enterprise.
  • OpenAI Codex: one ChatGPT sign-in across terminal, editor, cloud and app. Included in ChatGPT plans (limited on Free and Go) or usable with an API key.
  • Claude Code: terminal, IDE, desktop app and web, with detailed rules and automation settings, plus the Agent SDK to reuse its loop in your own app. Needs a Claude subscription or Console account.

Walkthrough: one task, three routes

Task: "Reject a blank email address on the bus-pass top-up form."

  • Copilot: open the issue on GitHub, assign it to Copilot, review the pull request.
  • Codex: describe it in the CLI or send it to Codex Cloud, review the diff.
  • Claude Code: ask in the terminal, approve the plan, review the commit.

Same goal, different route. Choose by where your team works and what your company has approved.

Where you see it at work

  • Banking and insurance: what's on the approved-tools list often decides first.
  • Product owners: issue-to-PR agents reward well-written tickets.
  • Business analysts: clear acceptance criteria help all three.

Try it today (10–15 min)

1. Pick one tiny task in a practice repo.
2. Try it in one tool; if you can, a second.
3. Compare: how many prompts, how much fixing, how easy to review?
Docs: https://docs.github.com/en/copilot · https://developers.openai.com/codex · https://code.claude.com/docs/en/overview

In one sentence

Copilot, Codex and Claude Code share the same core skills; pick by where your work happens and keep a human reviewing every change.

Which of the three have you tried, and what surprised you? 👇