Observation is becoming a crucial issue in the engineering of today's systems: the common practice for dealing with their complexity is to encapsulate their subcomponents abstracting away from their internal details, namely, focusing on their observable behaviour.

Starting from the framework for Observation within computer systems that we developed in (Viroli, Moro, Omicini, 2001), in this paper we study the impact of thinking about agents and multi-agent systems in terms of their observable behaviour.

Firstly, we apply our ontology for observation to agents. We claim that it can be significant to describe agents in terms of knowledge sources of such an ontology, that is, focusing on their observable behaviour.

Then, we proceed by extending the whole framework for observation so as to focus on the collaboration between observable systems, where an observation manifestation is also interpreted as an observation request for another system. This naturally leads us to model a multi-agent system in terms of agents sharing their knowledge through the continuous exchange of observation messages.