Hypertext environments such as the Web are rich with both word and
link cues that can be exploited by autonomous agents performing
distributed tasks on behalf of the user.  This paper characterizes
such environments and identifies the features that are most useful and
readily available.  We describe the adaptive representation of an
ecology of retrieval agents who attempt to capture important features
of their surroundings, and base their behaviors upon them.  We discuss
how such a representation allows the agents to interact with the
environments where they are situated.  Agents can internalize words
that are locally correlated with fitness, based on user feedback.
They are shown to outperform non-adaptive search by an order of
magnitude.  Furthermore, each agent learns new strategies at local
time and space scales, while the population evolves at a global scale.