well I understand that but let me clarify my problem.

suppose there is a cube with 10 unit long and 2 unit width. when the agent is walking along this cube, it must be way at least 2 units away, but if I use a circule, it walks at most 5 units away when it reaches the middle of the cube. I think it would make more problems with rectangular obstacles bigger this assumption.

yes, you don't have to decompose the cube the way i did, that's just generally the simplest solution.

to use your example in a better light, if you take a line from your avoider, to it's destination, and it intesects the barrier, then you need to get two points that create a perpindicular line from the obstacle, so, with the cube, you can fire a ray from the center of the obstacle, towards the perpendicular directions, and take those points as your obstacle radius, this gives you a variable sized circle based on the direction of the avoider, to the obstacle.

but as you can see, we still end up with a circle, regardless of how the object is decomposed, this is why the papers give solutions in terms of circles, the solution generally requires a point of some distance away from the obstacle in order to avoid collision, and since a circle defines such a shape, that's why it's used as the agent which is presented, since it's simple, and easy. it's up to you to select an algorithm which can accurately decompose the object.

if you look up the RVO2 library, and look at the source code, it has an algorithm for decomposing non-circle objects to generate the obstacle cone's.