Add random generation of two personalities with different parameters, or allow the personalities to independently choose the characteristics they need for a specific task. These aren’t even ordinary agents anymore; when two personalities compete with each other, working asynchronously on the same task but using completely different approaches, there might even be an additional mode that allows you to see their thought processes, meaning they’ll also be able to observe each other, spot mistakes, and monitor ongoing processes, thereby learning and refining their results. This can also be done with a single character; it’s just that different character traits will be generated for each task.
Parallel subagents can already approach parts of a task separately: https://learn.chatgpt.com/docs/agent-configuration/subagents. We'll share the idea for task-specific personalities and easier cross-review. We don't have a timeline for that mode.