Since the 90's, the Internet has tremendously evolved in terms of number and diversity of available services. In this trend, proxies are playing a central role and are spread all over the net. Today, the only functionality of proxies is to speed content access through caching only considering statistics based on use made by a user group, all together. Nowadays, evolutions are focusing on two main trends: specific infrastructures that do not comply anymore with the Web universality or individual applications like filtering. We propose to make proxies evolve from simple passive intermediaries to an open platform with advanced caching functionalities stressing on maintaining a high level of compatibility with existing paradigms. We ground our developments on two qualities already available in proxies, widely spread over the Net: the privileged position (plesiocentrism) within the network infrastructures and the perceptivity we extend to new dimensions, more specifically semantic (engnose). In a first step, we elaborate a new Web resources indexation binding to the documents a semantic space, based on the absolute position defined from the URL and a relative position defined with the links that connect those resources. We extend the usage statistics to those new dimensions to build up a topological space that take into account the localization of the documents and the user browsing into that space. We present algorithms and functionalities to build, maintain and take advantage of this topology. With the aim of accelerating web browsing, we use this statistical space to implement a prefetching system based on Markov's model. To extend the proxy perceptivity upstream to the users, we propose a mechanism to maintain a session, based on proxy-cookies. Therefore, we propose two new HTTP directives similar to those used for cookies. This paradigm also allows us to install personalized services with the support of the interaction concept and user profile. This last one allows us to tackle with mobility problems and to install proxies independently of physical network infrastructures. In a second move, we study how the proxy could take into account new dimensions in the semantic web and ontology context. The new technologies emerging like XML and annotations bring new information. That information can easily be processed by a computer system. As for the annotations, they considerably enrich the available informations in the proxy's perception through the classification or resources in ontology. We underline the opportunity for proxys, regarding their situation, to integrate functionalities of annotation server. We demonstrate what can be acquired that way and advantages to be won with the engnose definition as a new proxy quality that become perceptive to web disseminated knowledge. We present a new cache management based on virtual multi-level cache. We present an algorithm able to switch automatically to the correct domain regarding the ontological value of currently visited resources. Last, to demonstrate the validity of our propositions, we define the I3 platform (Intelligent Interactive Intermediaries): an architecture that supports all the various mechanisms presented above, but preserving integration with a minimum of impact on existing infrastructures. We define the concept of proxlet that is an intermediary agent generalization and allow the development of new user services. We demonstrate the value of our proposition by presenting opportunities brought by our platform. We take advantage of the interaction allowed by the session concept to implement functionalities also made possible by an advanced cache model based on topology, semantic and ontology. This platform allows the implementation of services such as browsing help, mobility filtering and service integration.