Agents Helping Agents: Issues in Sharing How-to Knowledge

degrading gracefully (i.e., safely, comprehensibly, and helpfully) when it can’t complete a task. Moreover, it should be straight-forward to describe to the agent a new task to be performed (at least for a task having straight-forward specification) and to know that what was told was understood. When necessary, the agent should point out (in advance, preferably) when the new task may interact unfavorably with other ongoing tasks and work with the user to pin down how to handle such cases. Another key requirement is that the agent be general-purpose: it should be able to use all of the sensors and effectors available to it to solve a wide range of problems. An agent with these capabilities should be of enormous value in making humans’s use of computers easier and more cost effective.