The COSY architecture is a hybrid BDI-architecture that includes elements of both the PRS and IRMA, and was developed specifically for a multi-agent testbed called DASEDIS [Haddadi, 1994][Burmeister and Sundermeyer, 1992]. The architecture has five main components: (i) sensors; (ii) actuators; (iii) communications; (iv) cognition; and (v) intention. The first three components are straightforward: the sensors receive non-communicative perceptual input, the actuators allow the agent to perform non-communicative actions, and the communications component allows the agent to send messages. Of the remaining two components, the intention component contains `long-term goals, attitudes, responsibilities and the like ... the control elements taking part in the reasoning and decision-making of the cognition component' [Haddadi, 1994], and the cognition component is responsible for mediating between the intentions of the agent and its beliefs about the world, and choosing an appropriate action to perform. Within the cognition component is the knowledge base containing the agent's beliefs, and three procedural components: a script execution component, a protocol execution component, and a reasoning, deciding, and reacting component. A script is very much like a script in Schank's original sense: it is a stereotypical recipe or plan for achieving a goal. Protocols are stereotypical dialogues representing cooperation frameworks such as the contract net [Smith, 1980]. The reasoning, deciding and reacting component is perhaps the key component in COSY. It is made up of a number of other subsystems, and is structured rather like the PRS and IRMA (see above). An agenda is maintained, that contains a number of active scripts. These scripts may be invoked in a goal-driven fashion (to satisfy one of the agent's intentions), or a data-driven fashion (in response to the agent's current situation). A filter component chooses between competing scripts for execution.