Comment by xutopia

The article actually explains it better than I can but essentially you do not want the agent to get desperate which causes it to burn extra tokens. Planning phase doesn't test assumptions by going straight in the code and testing out parts of the ideas to execute so it has to consider way more in planning than it would if it could go quicker into the coding phase.