APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH E Edge: an arrow showing what may run next. Edge agent: an agent that runs on local hardware, such as a phone or small device, instead of in a distant data center. Edge case: an unusual but possible situation that tests whether a system copes with the odd corners of its job. Egress: traffic leaving a system. Egress control: limits on outgoing network traffic. Embedding: a list of numbers that captures the meaning of a piece of text so that similar meanings sit close together. Endpoint: a specific web address that accepts a certain kind of request. Environment: a separate copy for a stage such as development, staging, or production. Environment variable: a named setting stored outside the code, often used to hold secrets such as keys. Episodic memory: memory of specific past events, like remembering what happened in an earlier run. Escalation: retrying with a stronger model when a cheaper one falls short. Escalation path: the planned route by which a problem is handed up to a human or a more capable system. ETCLOVG: the name of the seven-layer framework in this book, standing for Execution, Tooling, Context, Lifecycle, Observability, Verification, and Governance. Eval: a structured test that measures how well an AI system does its job. Evaluation: an organized test that measures how well an agent performs on a set of tasks. Evaluation harness: the system that runs tests on an agent and measures its quality. Evaluation report: a written summary of test results, usually including success rate, cost, and speed. Evaluation set: a fixed collection of test tasks with known good answers. Evaluation suite: a set of test cases used to check quality. Evaluation-driven iteration: improving a system step by step, guided by measured test results instead of guesses. Eviction: removing items from the window to make space. Exact match: checking that output is identical to the expected answer. Exact-match scoring: grading in which an answer counts as correct only if it matches the expected answer exactly. Execution layer: the first ETCLOVG layer, covering the loop that drives the agent. Executive summary: a short overview of the main points at the start of a report. Exponential backoff: a retry method that doubles the waiting time after each failed attempt. External memory: information stored outside the window that the agent can read back. External notes: information the agent writes down outside its context window, such as in a file, so it can look it up later.