I would like to add a point that I think is very important: In Long-Horizon agent or task, Verifier > Planner, because long tasks continue to accumulate and summarize context, it will make related work very challenging.
Some iron laws of a certain stage may be lost in subsequent stages and no longer valid.
We all know the butterfly effect.
There will be many butterfly effects in long tasks.
An inadvertent and inaccurate decision in the early stage will produce a lot of debris and friction in the task execution process after a few hours without intervention.
I attached great importance to the capabilities of the planner during the architecture design phase and spent a lot of energy on carving.
However, after the project was actually run, I found that the verifier has assumed a more core function.
It is an important means to prevent the butterfly effect and similar problems.
Its role is actually greater than that of the planner.
It has to integrate a lot of states, target tracking, current effect comparison, correction strategies and other very important functions.
It is strongly recommended to spend the same or more energy on the verifier during the design phase.
