This paper presents a new methodology for building special-purpose software agents that capture and access information in large, heterogeneous, distributed information environments. It allows rapid prototyping of information agents for solving a wide range of retrieval tasks with guarantees on performance. The key idea is to exploit underlying structure at various levels of granularity to build partial models that act as high-level indices with task-specific interpretations. These partial models are constructed using modules called navigators. Information agents are configured by using effective communication protocols to connect structure detectors and navigators. This methodology is applied to the design and implementation of information agents in two contexts: one for retrieving stock market information from scanned copies of newspapers and another for retrieving technical reports from the Internet.