How to set reasoning effort and subagent instructions for GPT-6.2 in Codex
OpenAI added a prompting guide for GPT-6.2 Astra and Sol to the Codex docs. Its main advice is to match reasoning effort to the depth of the work and to tell Codex up front that it may hand parallel work to subagents.
- Pick reasoning effort by how deep the investigation goes. The guide suggests Medium for a quick survey, High for tracing one suspected cause across several systems, and Extra High when several causes may interact. In the CLI, change it with
/model; put the value you always want in the config file.
# ~/.codex/config.toml
model = "gpt-6.2-sol"
model_reasoning_effort = "medium"
The configuration reference lists low, medium, high, xhigh, max and ultra. Which levels you can actually pick depends on the model and client.
- When a task has several parts that can move ahead independently, add "use subagents as needed" to the request. Codex decides when and how to split the work. If the same kind of work comes up often, put that instruction in the relevant skill or AGENTS.md and say when it applies.
## Research requests
- When there are several independent questions, use subagents as needed.
With Ultra, Codex brings in subagents on its own. The guide says to save it for ambitious work that splits cleanly into several questions that each need a deep investigation.
- Keep AGENTS.md lean. Keep only requirements and conventions Codex can't infer from the code, tests or docs, plus any workflow the project depends on, and leave routine decisions about how to do the work to Codex.
Also tell Codex what you'll look for when you review the result. The guide's example: if an icon needs to be recognizable in a 24-pixel toolbar, say so.
Sources
Get notified when a new tip lands
Tip alerts only go out in the app. Install it and turn notifications on.