Manus creates a todo file when giving an agent a complex task, then checks items off one by one as progress is made. That apparently reduces the lost-in-the-middle problem and keeps the work aligned with the goal.
Many AI agents, including Claude Code, use this method, and the more complex and lengthy the task is, the more important it becomes not to lose the thread. But really, isn't this exactly what humans do when they work well too?
https://harvest.pub/shared/8f21578a-bf57-4ce4-8785-ec57cb21e19e?highlight=1753491804488
