Codex vs Claude Code vs your own loop: pick the runtime per agent
There are three honest ways to run an AI agent today, and most platforms quietly pick one and pretend the others don't exist. You can run OpenAI's Codex — the actual CLI, with its own tool loop, its own context compaction, its own way of working a repository. You can run Anthropic's Claude Code — same idea, different strengths, Anthropic models only. Or you can run a purpose-built loop: your own orchestration around any model API, which is more work but gives you control the packaged agents can't — over the prompt, the tool set, the permission checks on every call.
The uncomfortable truth is that the right answer changes task by task. A repository migration wants Codex's repo instincts. A long writing-and-research job wants a loop you can instrument. A build task on Anthropic models wants Claude Code. Choosing once, for a whole product, means being wrong most of the time.
Why imitating them is worse than running them
A tempting shortcut is one generic loop with a prompt that says "act like Codex." It doesn't work, because what makes these agents good isn't their persona — it's their machinery. Codex's compaction strategy, its command execution, its file editing behavior are all in the binary, not the prompt. So Choreo runs the real binaries, driven over their real protocols, each in an isolated workspace. When an agent is set to the Codex runtime, the thing executing is Codex.
Where the purpose-built loop earns its place
Our own loop — we call the prompt side of it Planx — exists for the things a packaged agent can't give you:
- Any model. Anthropic, OpenAI, Google, or any OpenAI-compatible endpoint you bring — the packaged agents are each built around one vendor's models.
- A compiled prompt. The system prompt is assembled per turn from instruction modules chosen by what the task actually looks like, instead of one prompt trying to cover everything.
- Permission checks on every tool call. Every capability — file writes, email, shell, phone calls — resolves to allow, ask or deny before it runs, and delegated sub-agents can only ever get a narrower set than their parent.
The part that matters: pick per agent
The runtime is a per-agent setting, not a project setting. One project can hold a lead agent on our loop that delegates a repo task to a Codex agent and a review task to a Claude Code agent — in the same run, each on its own model, each with its own limits. That's the actual shape of the work: teams are mixed, so the engine should be too. The full breakdown of what each runtime supports lives on the runtimes page — or skip the reading and put a Codex agent and a Claude Code agent in the same team yourself, free, no card.
See it for yourself — free, $2 of credits once you confirm your email, no card.
Start free