Fine grained context per task instead of per project.

What variant of Codex are you using?

IDE Extension, CLI

What feature would you like to see?

Currently all AI agents are architected on a per project basis. This causes them to make exploration of the project a large and time-consuming task. If users could feed agents with more specific per task context a lot of time and tokens would be saved per task. Example, have several setups per project with agents.md + mcp + files + prompt (or mcp + files + prompt) for each specific task. I could select one of those prepared contexts (example with a dropdown), add the prompt and run the task. The result on average less exploration, less tokens and less time = faster development speed.

Additional information

https://github.com/anthropics/claude-code/issues/20292#event-23168262041