4 in each different world situation. Mediation remains active at all times and is largely " transparent " to the layers: each layer acts as if it alone were controlling the agent, remaining largely unaware of any " interference " — either by other layers or by the rules of the control framework — with its own inputs and outputs. The overall control framework thus embodies a real-time opportunistic scheduling regime which, while striving to service the agent's high-level tasks (e.g. planning, causal modelling, counterfactual reasoning) is sensitive also to its low-level, high-priority behaviours such as avoiding collisions with other agents or obstacles. Through a number of single-and multi-agent coordination experiments addressing such issues as the role of prediction in resolving inter-agent goal conflicts, variability in levels of agent sensitivity to environmental change, and the production of emergent behavioural patterns, the TouringMachine architecture has been shown to be feasible and that, when suitably configured, can endow rational autonomous agents with appropriate levels of effective, robust, and flexible control for successfully carrying out multiple goals while simultaneously dealing with a number of dynamic multi-agent events. The integration of a number of traditionally expensive deliberative reasoning mechanisms (for example, causal modelling and hierarchical planning) with reactive or behaviour-based mechanisms is a challenge which has been addressed in the Touring-Machine architecture. Additional challenges such as enabling effective agent operation under real-time constraints and with bounded computational resources have also been addressed. The result is a novel architectural design which can successfully produce a range of useful behaviours required of sophisticated autonomous agents embedded in complex environments. [2] Rodney A. Brooks. A robust layered control system for a mobile robot. 3 suited to purely deliberative agents. What is most likely, however, is that the majority of real-world domains will require that intelligent autonomous agents be capable of a wide range of behaviours, including some basic non-deliberative ones such as perception-driven reaction, but also including more complex deliberative ones such as flexible task planning, strategic decision-making, complex (e.g. time dependent, prioritized) goal handling, or predictive reasoning about the beliefs and intentions of other agents. 3 TouringMachines: a hybrid solution My position is that is it both desirable and feasible to combine deliberative and non-deliberative control functions to obtain effective, robust, and flexible behaviour from autonomous, resource-bounded task-achieving agents operating in real-time multi-agent environments. In particular, the research highlighted here is concerned with the design and …