Which Agent Should I Start With?¶
Ralph Workflow is the operating system for autonomous coding — a free and open-source CLI that runs the coding agents you already use on your own machine.
It is for developers and technical teams with work that is too big to babysit and too risky to trust blindly.
What makes it different is the handoff: Ralph Workflow is built to bring back a reviewable result in your repo instead of just a transcript and a claim that the task is done.
Why use it now? Because you do not need to switch your whole toolchain first. Pick one agent you already trust, run one real backlog task tonight, and decide tomorrow whether the result is something you would actually merge.
The short answer¶
For a first Ralph Workflow run, start with the agent that is already installed, already authenticated, and already familiar on your machine.
Do not optimize this choice too hard.
The main first-run question is not “which model is theoretically best?” It is:
Can I get one real unattended run to finish with a reviewable result?
If one agent is already working for you today, that is usually the right first choice.
Fast decision guide¶
Start with Claude Code if…¶
claudealready works on your machineyou want the most straightforward documented path
you care most about planning quality and clean review handoff
Why this is a good first fit:
strong default choice for end-to-end unattended work
commonly the clearest first-run path for Ralph Workflow users
good when you want to judge the workflow, not compare providers yet
Start with Codex if…¶
you are already using OpenAI tooling
you want an OpenAI-first setup
you expect to care about cost control or provider familiarity more than picking one “best” agent
Why this is a good first fit:
strong option for teams already standardized on OpenAI
solid review and implementation choice
keeps the first run inside tools you already know
Start with OpenCode if…¶
opencodealready works on your machineyou want multi-provider flexibility from the start
you already have an OpenCode setup you like and do not want to reconfigure everything just to try Ralph Workflow
Why this is a good first fit:
preserves your existing gateway setup
lets Ralph Workflow orchestrate the agent stack you already use
good for teams that switch models often
Best first-run rule¶
Pick the path with the least setup friction.
That usually means:
the agent is already installed
the agent is already authenticated
you have already used it manually in the same environment
If two agents are equally ready, prefer the one you would be happiest reviewing output from tomorrow morning.
What matters more than the agent choice¶
For a first run, these matter more than whether you picked Claude Code, Codex, or OpenCode:
choosing a small real backlog task
writing a clear one-paragraph spec in
PROMPT.mdmaking sure the repo is safe to test in
judging the result with one question: would I merge this?
A good first task will teach you more than switching agents three times.
What not to do¶
Avoid these first-run traps:
picking an agent you have not installed yet if another one is already working
trying to compare multiple providers before you have seen one successful handoff
choosing a broad risky task just because the model feels strong
treating the first run like a benchmark instead of an honest workflow test
If you do not have any agent set up yet¶
Ralph Workflow does not replace the coding agent itself.
Before your first run, install and authenticate one supported agent CLI first. Then come back and use Ralph Workflow to orchestrate it.
If you want the shortest honest path after that, continue with:
Recommended first-run sequence¶
Pick the agent that is already working on your machine.
Use Getting Started for the fastest real-task path.
If
PROMPT.mdis blank, use First-Task Prompt Templates.Run one bounded backlog task.
Review the result tomorrow and ask: would I merge this?
That is enough to tell you whether Ralph Workflow is useful in your real environment.