How to Get More From Codex With Models and Loops

TL;DR

Use GPT-5.6 Soul for the hardest problems and Luna for easier work, then extend Codex with delegated threads, a concise AGENTS.md file, plugins, browser control, skills, and persistent goals. The strongest workflow assigns each task an appropriate model and effort level, while loops repeatedly test, diagnose, fix, and retest until a measurable target is reached.

Transcript

There are so many unknown tips and tricks to get the most out of Codex and even advanced users often times don't know all of them. That's what this video is about. I put together everything I have learned about Codex after using it for over a thousand hours and it doesn't matter if you're a beginner or you're very experienced, you will get somethin... Read More

Key Insights

• GPT-5.6 Soul is the recommended model for the hardest problems, while Luna is recommended for most other work. The presenter typically uses Soul with high effort and Luna with high or extra-high effort, while generally avoiding Soul at the light effort setting.
• Luna Max can provide a better score at a lower budget than Terra High according to the graph presented. This comparison supports a simplified selection rule: choose Soul when maximum capability matters, and choose Luna when the task does not require the largest model.
• Fast mode is stated to cost two and a half times as much while delivering a one and a half times speed increase. The presenter therefore recommends regular mode when time is available, while leaving fast mode as an optional choice when speed matters more.
• Project-scoped threads can be created from an existing Codex thread with a specified model and thinking effort. This allows a main thread to delegate tasks to another model configuration while keeping that work inside the same project folder.
• Codex threads can see other threads, enabling one thread to coordinate or monitor several parallel tasks. A supervisory thread can watch whether the others are progressing and prompt a stopped thread to continue, creating a managed multi-thread workflow.
• AGENTS.md rules are loaded frequently into the context of agents and can accumulate instructions designed for older models. Reviewing the file after a new model release can identify stale, unnecessary, or counterproductive rules that should be revised or removed.
• Codex browser control can perform real-world interface tasks, not only test websites or applications. The demonstration creates an archive folder and moves two documents into it, while the presenter also reports using browser control for tasks involving DNS records and Supabase scaling.
• Codex goals are persistent loops that continue working toward a measurable outcome. A goal can run a benchmark, examine failures, modify the project, and rerun the benchmark repeatedly until it reaches a specified threshold, with some reported runs continuing for up to 12 hours.

Questions & Answers

Q: Which GPT-5.6 model should I use in Codex?

Use GPT-5.6 Soul for the hardest problems and GPT-5.6 Luna for easier or routine work. The presenter usually pairs Soul with high effort, occasionally dropping to medium, and uses Luna with high or extra-high effort. The graph shown also indicates that Luna Max can score higher while costing less than Terra High, making Terra less attractive in the compared configurations.

Q: Should I enable fast mode when using Codex?

Fast mode is optional and is most appropriate when saving time is worth the additional expense. According to the transcript, fast mode is two and a half times as expensive but provides a one and a half times speed increase. The presenter recommends leaving it off when time is available because the proportional cost increase is greater than the proportional speed improvement.

Q: How do delegated threads work in Codex?

A Codex thread can create another project-scoped thread and specify the model, thinking effort, and prompt for that delegated task. The new thread works in the same project folder as the original thread. This makes it possible for a primary thread to manage the overall job while assigning focused work to another model configuration suited to that task.

Q: How can one Codex thread monitor parallel tasks?

Codex threads can see other threads, so a dedicated supervisory thread can monitor several tasks running in parallel. The presenter suggests using a main thread to check whether each worker is still running and progressing appropriately. If a worker stops, the monitoring thread can prompt it again and push it forward, allowing centralized oversight of concurrent work.

Q: Why should I review my AGENTS.md file after a model update?

An AGENTS.md file can accumulate rules created to compensate for the behavior of earlier Codex models. Those rules may become stale when a new model is released, yet they continue to be loaded frequently into agent context. Reviewing the file can uncover instructions that should be revised or removed, reducing unnecessary context and keeping current workflows aligned with the newer model.

Q: What can plugins add to Codex?

Plugins can connect Codex to additional services and functions. The featured example is the Zapier MCP, which the presenter says provides access to more than 9,000 applications, including Gmail, Trello, Asana, and Google Docs. The setup path shown goes through Settings, Integrations, Plugins, and MCPs, where an installed Zapier plugin can be selected and used.

Q: What tasks can the Codex browser perform?

The Codex browser can handle interface-based work beyond testing a website or application. In the demonstration, it creates a new archive folder and moves two selected documents into it from a short prompt. The presenter also reports using it for more complex operations, including changing DNS records and automatically scaling a Supabase instance up and down.

Q: How do Codex goals improve a project automatically?

Codex goals keep the agent working toward an overarching result instead of returning only one response to one prompt. The example instructs Codex to run a private benchmark, inspect how the project failed, attempt fixes, and repeat until the score reaches at least 90%. The presenter reports that such loops can run for hours, including some sessions lasting up to 12 hours.

Summary & Key Takeaways

• Effective Codex use begins with matching model cost and capability to the task. The presenter recommends GPT-5.6 Soul at high effort for difficult problems and Luna at high or extra-high effort for easier work. Fast mode offers greater speed, but its stated cost increase is larger than its stated performance gain.

• Threads can delegate work to selected models and thinking budgets while a primary thread coordinates progress. Project-scoped threads share the same project folder and can see other threads, enabling parallel execution and supervision. A monitoring thread can watch active tasks, detect when one stops, and prompt it to continue working.

• Codex becomes more capable through maintained instructions, plugins, browser use, reusable skills, and persistent goals. Old AGENTS.md rules should be reviewed after model releases. Goals can repeatedly run benchmarks, investigate failures, apply fixes, and retest until a defined threshold, such as a 90% benchmark score, is achieved.