multi-step planning

Planning is the model's ability to look at a goal and lay out a sequence of actions that gets there, rather than reacting one move at a time. Ask it to organise a trip, refactor a codebase, or research a question, and a capable model can decompose the goal into subtasks, order them sensibly, anticipate what each step needs from the ones before, and adjust when something does not go as expected. It is the difference between answering a question and carrying out a project.

This capability is what makes agents possible: a planning model can decide to search, then read, then write code, then check, looping until the goal is met. But planning is among the harder things to do reliably, because errors compound. A flawed early step poisons everything downstream, and a long plan offers many places to drift off course or forget the original aim. Models plan best when they can act, observe the real result, and replan against it, rather than committing to a long blueprint up front and following it blindly.