Split planning and coding across two Codex agents
They routed planning to GPT-5.5 and most coding to a cheaper GPT-5.4 Mini agent.
- What they did
- They said they did not use Codex’s built-in agent flow. Instead, they set up their own two-step system in which GPT-5.5 first outlined the work, then a GPT-5.4 Mini agent handled the implementation, including editing files.
- How it went
- The point was to keep routine coding work cheaper by reserving the larger model for planning. They did not share timing, savings figures, or concrete examples, and there were no reported failures.
- Tips
- If you try this, make the planner produce small, checkable tasks and have it review each step from the executor. Keep a human approval pause before any irreversible action, such as submitting code or deleting files.
Try it yourself with Codex
In [repo path], help me set up a two-agent coding workflow for [task]. First, use a stronger model to make a short plan with clear subtasks and checks. Then send each subtask to a smaller, cheaper agent that can edit files and implement the work. After each step, have the planner review the result and decide whether it needs another pass. Pause and ask for my approval before any irreversible action, including submitting changes, deleting files, or making payments. When everything passes the checks, give me the final plan, the code changes, and any parts that needed rework.
Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.
Get 3 agent uses like this every morning
With the day's top AI news. Free, and one click to unsubscribe.
Related agent uses
Turn GitHub issues into draft PRs with a long-running agent
A developer used Dots to read GitHub issues and produce rough pull request drafts, mostly as a starting point for later work.
Use a coding agent to fix a Slack-reported bug
A Codex agent handled a Slack issue on its own, and an ops agent later confirmed the fix in the same thread.
Use Codex as a first-pass code submission filter
A commenter used Codex with a custom agents.md to check candidate code and automatically turn away submissions that broke simple file-change limits.
Ask Codex for five admin screen redesigns
A developer used Codex CLI to create five alternative admin-screen designs and then chose the one they preferred.
Set up a separate sandbox QA pass
Codex turned an existing Claude login into a second reviewer that checked an AI system without changing the user’s credentials.
Flag coding conflicts before agents edit
They set up a local layer that makes coding agents announce planned changes so overlapping architecture work gets caught early.