Abstract. Most expert system control strategies focus on a narrow view of problem solving, with the resulting expert systems failing to accomplish a good interaction with the task environment during problem solving. Known as the pregnant man problem, such a deficient external behavior is typical of many expert systems that ask absurd questions of the user. Right interaction with the task environment has been already addressed as a verification problem. A new framework for realizing a good external behavior is presented in this paper, that consists of incorporating constraints on external behavior in the control strategy of the expert system. Called input-driven control, it cooperates with the standard control strategies of the Milord II expert system shell for realizing both a correct problem-solving behavior and an appropriate external behavior.