APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH C Cache: a temporary storage spot that keeps a copy of something so it can be reused quickly instead of being produced again. Calibration: comparing a judge against human scores to check trustworthiness. Canary release: a careful way to launch a change by first giving it to a tiny group of users and watching for trouble. Capability attestation: a trustworthy statement from a tool or server about what it can and cannot do, which some protocols do not yet require. Capstone assessment: a final exercise that tests several skills together. Case type: a label on a test item saying whether it is a normal, edge, or adversarial case. Central Processing Unit (CPU): the main chip in a computer that carries out instructions and does the general thinking work. Changelog: a list of changes between versions of a tool. Checkpoint: a saved snapshot of an agent's progress that it can return to later if something goes wrong. Chunking: cutting long documents into smaller pieces so that the right piece can be found and used later. Citation: a pointer from a claim to its supporting source. Cleanup: closing and deleting resources a run no longer needs. CLI: a program used by typing text commands. Client: the side of a connection that asks for something, such as an agent asking a tool server to do a job. Closed model: a model whose inner workings are kept private and that you can only use through the company that hosts it. Closed, hosted model: a model reachable only through its provider's API. Coding agent: an agent whose main job is reading and writing programs. Cold storage: the place for old, full records that are kept safe but rarely needed, so they are not loaded into active use. Cold tier: complete transcripts and raw data in long-term storage. Command injection: user input placed into a command that runs. Compaction: summarizing older parts of a conversation into a shorter form to free up space in the context window. Compensating action: a new action that cancels an earlier action's effect. Completion: the text that a model writes back in response to a prompt. Completion detection: harness logic that decides whether criteria were met. Compliance: following the laws, regulations, and organizational rules that apply. Component: a distinct piece of software with one clear job. Component map: a diagram or list showing the parts of a harness and how information flows between them. Composable: designed in small pieces that can be combined in many ways instead of one giant all-in-one piece. Computer-use agent: an agent that operates a computer by looking at the screen and simulating mouse and keyboard actions. Concurrency: many things at once. Conditional: an if-this-then-that rule. Configuration: the set of adjustable choices that define how a harness behaves. Container: a lightweight, sealed box on a computer that holds a program and its needed files, separate from everything else. Containment: stopping damage from spreading. Contamination: benchmark content leaking into training data. Context: all the information placed in front of the model at the moment it produces an answer. Context bloat: a situation in which the context window fills up with too much low-value text. Context builder: the harness component that decides what goes into the context window for each step. Context engineering: the skill of choosing and arranging all the information a model sees across many turns. Context isolation: giving each agent only the information it needs, so it is not distracted by unrelated material. Context rot: the drop in a model's accuracy that happens as its context window grows very large. Context window: the limited amount of text a model can look at all at once. Continuation hook: a hook that reminds the agent of its goal so that it does not quit before the work is finished. Continuous evaluation: running quality tests automatically and regularly as part of building and releasing software. Control flow: the order in which steps happen and which path is taken at a choice. Control plane: the layer of a system that decides how agents run, talk to each other, and recover from trouble. Convention: a community habit that is not formally enforced. Conversation hijacking: an attack in which outside text takes over the direction of an ongoing conversation with an agent. Coordinates: numbers giving a position on a screen. Cost ledger: a running count of what each tier has consumed. Cost per task: the average amount of money spent to complete one job from start to finish. Coupling: how tightly two parts depend on each other. CPU: the short form of Central Processing Unit (CPU). Covert tool invocation: an attack in which a tool is secretly triggered without the user noticing. Cross-session continuity: the ability to remember information across separate runs instead of starting from nothing each time. Crew-style framework: a toolkit for building several agents that each play a named role, like members of a team. Curated dataset: a deliberately chosen and maintained set of test cases. Curated list: a chosen and organized collection of links to good resources. Cycle: a loop in which a step leads back to an earlier step.