APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH A A/B testing: comparing two versions of something by running both on the same tasks and seeing which one performs better. Access control: deciding who or what may do which things. Accountability: a named party responsible for results. Accuracy: how often a system gives the correct answer, usually shown as a percentage of all attempts. Action tool: a tool that changes something in the outside world, such as sending an email or deleting a record. Active context: the text currently sitting inside the model's context window, as opposed to older material stored elsewhere. Adversarial case: a test built to trick or attack the system on purpose, such as a resume that hides a malicious instruction. Agent: a software system built around a model that can take actions on its own to pursue a goal. Agent harness: another name for the harness, meaning all the software wrapped around a model so it can work as an agent. Agent instructions file: a plain text file kept at the top of a project folder that tells an agent how the project works, such as how to build it and what to avoid. Agent loop: the repeating cycle in which an agent thinks, takes an action, looks at the result, and then repeats. Agent Operations (AgentOps): the practice of running AI agents reliably in real use, similar to how Development Operations looks after ordinary software. Agent2Agent (A2A): a protocol for agents to talk to other agents. AgentOps: the short form of Agent Operations (AgentOps). Agent-to-agent communication: the agreed ways that separate agents send messages, tasks, and results to one another. Agentic: describing a system that works in a loop of deciding and acting instead of answering a question once and stopping. Agent-to-agent standard: an agreed format for independent agents to cooperate. AGI: the hypothetical ability to learn and do almost any intellectual task a human can. AI: the short form of Artificial Intelligence (AI). Alert: an automatic warning sent to people when a measured value crosses a limit that someone chose in advance. Alert fatigue: ignoring alerts because there are too many. Allow-list: a list naming the only tools or actions that are permitted, with everything not on the list blocked. Anchor: an item that is never dropped, such as the original task. API: the short form of Application Programming Interface (API). API key: a secret string of characters that proves to a service that a request comes from an authorized user. Application Programming Interface (API): a doorway that lets one piece of software ask another piece of software to do something, such as calling a model. Approval fatigue: people approving without reading because of too many prompts. Approval gate: a checkpoint where a human must approve an action before it proceeds. Approval hook: a hook that stops the agent and waits for a human to say yes or no before a risky action goes ahead. Architecture: the high-level arrangement of a system. Artificial General Intelligence: see General AI. Artificial Intelligence (AI): computer systems that perform tasks that normally seem to need human thinking, such as understanding language or recognizing images. Attack surface: the total set of places where an attacker could try to get into or manipulate a system. Audit: an independent review of what a system did. Audit log: a permanent record of who or what did which action and when, kept so it can be reviewed later. Audit trail: a tamper-resistant record of what happened. Auditability: the quality of a system that lets an outsider check afterward exactly what it did and why. Auditing: reviewing records after the fact to confirm proper conduct. Authentication: proving identity. Authorization: deciding what an identity may do. Automated harness optimization: tuning the harness through automated experiments. Autonomous context compression: a method in which the agent itself decides when and how to shrink its own history to save space. Autonomy: freedom to act without approval at each step.