This research is partially funded by the ESPRIT Programme of the Commission of the European Communities under projects number 5248 (KADS-II) and 5477 (CONSTRUCT). The partners in KADS-II are Cap Gemini Innovation (F), Cap Gemini Programator (S), Netherlands Energy Research Foundation ECN (NL), Eritel SA (ESP), IBM France (F), Lloyd's Register (UK), Swedisch Institute of Computer Science (S), Siemens AG (D), Touche Ross MC (UK), University of Amsterdam (NL) and Vrije Universiteit Brussel (B). The partners in CONSTRUCT were Siscog (P), Renault (F) and Vrije Universiteit Brussel (B). This research is also financed by a ``Inter-Universitaire Attractie Pool'' on knowledge system and the IMPULS project ADIOS, both from the Belgian government.

The knowledge level has broader scope than knowledge systems alone. It can be used for characterising the behaviour of any complex system, for example of an autonomous agent, even if it is behaviour- rather than knowledge based. In addition the knowledge level can be used to describe different aspects of intelligent behaviour, like learning [41][17]. In this paper I only consider the problem solving or performance aspects.

An alternative to two step rationality is to model an agent as consisting of a structure of smaller agents, each one being specialised and perfectly rational in its specialisation [37]. I prefer to reserve the notion of agent for an embodied and observable being. Its application as a structuring device makes it counterintuitive to account for the dynamic configuration of knowledge in different models.

One could consider the rationalisation of configuration. Such a rationalisation would need to refer to detailed knowledge of the agent's constitution, evolution and adaptation over time, and details of the embedding and the interaction with the environment. This may be possible but would need to refer to details of the agent's physical embedding and adaptation to the world. This does not seem very practical and uncompatible with a black-box treatment of the agent. Much of this knowledge should therefore be considered tacit, i.e., implicit in the structure and interaction of the agent-environment system.

The action of hitting a nail with a hammer can be assumed to entail that the nail is deeper in the wood, that the nail is deeper in the wood with some probability, or that a loud sound has been produced. The 'effect' of the action simply depends on what assumption the agent makes about it.