In this article, we would like to detect boundaries of objects with the help of a multiagent system made up of reactive agents. The reactivity being very important, the agents' behavior is very simple (perception-action): they have the capacity, nevertheless, to adapt locally to what they consider their environment, that is to say the image. An agent can move and has its own position in its environment. The basic behavior for an agent consists of following the highest brightness gradient and inscribing its path, if estimating to be on an edge, in all the agents' shared memory. Its path thus corresponds to edges which are found in the image. Please note that, in order to be noise resistant, the path is actually stored in the shared memory only if it is long enough. The notion of shared memory is very important because it allows the interaction among agents and the coordination of their actions. The agents actually use already found edges for finding new ones or complete those already detected. We have tested this system on different gray scale images scenes, but as well on synthetic scenes allowing analysis of thus obtained results. The results are promising and especially fast. Our multiagent system has been tested on a single-processor computer, and it has been noted that the number of agents in a simulation neither affects the quality of the result nor CPU time necessary for segmentation of a given scene. We think that this approach is original in its use of agents and may be used to implement parallel image processing by assigning, for instance, an agent to each processor.