APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH I Idempotency: the property that a tool can be run more than once with the same input and still leave things in the same final state. Idempotency key: a unique label attached to a request so the system can recognize and ignore an accidental repeat. Idempotent: safe to call more than once with the same effect. Identity handling: the safe management of who or what is making a request and what they are allowed to do. Important flag: a mark on a message meaning it must never be dropped. Incident: an unplanned event in which a system fails or behaves harmfully. Incident commander: person coordinating a response. Indirect injection: a prompt injection that arrives hidden inside content the agent fetches, such as a web page or a file. Indirect prompt injection: harmful instructions hidden in content the agent reads. Inference: actually running a model to get an answer from it. Infinite loop: a repetition that never ends. Input schema: the agreed structure of the information a tool expects to receive. Input schema and output schema: the agreed shape of a tool's request and reply. Input tokens: tokens sent to the model. Input-output logging: recording only the request and the final answer. Instruction hierarchy: a ranking that says whose instructions win when they conflict, such as the developer's rules over text found in a document. Instruction manager: the harness component that assembles the agent's standing rules into one set of instructions. Integration test: a test that checks whether several parts of a system work correctly together. Interceptor: a hook that can inspect, change, or stop what passes through. Interoperability: the ability of different systems to work together smoothly because they follow the same standard. Irreversible action: an action that cannot easily be undone. Iteration: one pass through a cycle, or one round of improvement in a repeated process.