Sensing, in the context of an AI system, is the act of inspecting the state of external systems for the purpose of evaluation. It is done by the Agent Harness. Sensing is distinct from Retrieval or Navigation. While Retrieval might use a bash command to cat a file, and Navigation might use a bash command to ls a directory structure, sensing is about getting state as the first part of a Feedback Loop. It is sensing if some reasoning is done based on the state returned.
Sensing is important in the development of more autonomous and effective Agentic workflows, as it can give an Agent “insight” into its environment and its effects on it. This forms the basis for Feedback Loops.
Examples of useful “senses”:
- Reading a log file
- Running a build and test loop and getting the results
- Running a bash command that checks if expected files were created