THE VERDICT
● Claude Code for interactive work. Still first among harnesses for long sessions.
● Codex if you want the same agent across CLI, IDE, web and desktop.
● GitHub Agentic Workflows for repeatable tasks that should run without you.
● The thing to notice: CI means unattended, and unattended is where agents go wrong.
Where each one runs
| Claude Code | Codex | GH Agentic Workflows |
|---|
| Runs in | Terminal, IDE, desktop | CLI, IDE, web, desktop | GitHub Actions |
| You are watching | Usually | Usually | No |
| Metering | Weekly pool, shared | Five-hour cap | Actions minutes |
| Credentials available | Yours, locally | Yours | Repo secrets |
| Status | GA | GA | Technical preview |
THE CREDENTIALS ROW IS THE ONE TO READ TWICE
An agent in CI has whatever your Actions secrets hold — deploy keys, registry tokens, cloud credentials. That is usually more than an agent in your terminal has.
METR traced 1,200 agents through a package repository nobody was monitoring. CI is exactly that shape of environment, and it runs while nobody is looking.
What each is genuinely good for
- Claude Code — long interactive sessions on a codebase you know. Ranks first among harnesses for a reason, and the weekly pool now sits lower after 14 September.
- Codex — one agent across four surfaces, metered separately from chat, and GPT-6 Astra included on Plus.
- GitHub Agentic Workflows — the repeatable things nobody wants to do. Dependency bumps, test triage, mechanical refactors across many files.
If you use the GitHub one
- Scope the secrets it can see. A workflow that opens pull requests does not need deploy credentials.
- Require review on anything it produces. Research on agent pull requests found larger diffs merge less often — cap the size.
- Set a minutes budget per workflow, not per month. A loop exhausts a monthly budget in an afternoon.
- Log what left, not just what ran. CI has network access by default.
- It is a technical preview. Do not put it on the path to production yet.
Which one
| If you want to... | Use |
|---|
| Work through a hard problem with the agent | Claude Code |
| Same agent everywhere you work | Codex |
| Automate a boring repeated task | GitHub Agentic Workflows, with scoped secrets |
| Run anything near production | Not the preview. Wait for GA |
FAQ
What are GitHub Agentic Workflows?
Coding agents running inside GitHub Actions, announced in technical preview. That means they run in CI rather than on your machine.
Is it safer than running an agent locally?
Different, not safer. It is contained from your laptop but has access to repository secrets and runs unattended, which is the harder case.
Does it replace Claude Code?
No. Interactive work and CI automation are different jobs, and most teams will end up doing both.