APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH M Machine learning: improving at a task by studying examples. Machine Learning Operations (MLOps): the practice of building, releasing, and running machine learning systems reliably. Machine-learning engineer: a person who trains, tunes, and measures models. Managed service: a product another company runs for you in return for payment. Matched-pair test: comparing inputs identical except for one detail. Maximum-iteration safeguard: a limit on loop turns so a loop cannot run forever. MCP: the short form of Model Context Protocol (MCP). Max iterations: a setting that caps how many times a loop may repeat, to prevent endless running. MCP client and server: the program that uses tools and the program that offers them. Memory: any way a system keeps information so it can use it later. Memory hierarchy: a layered arrangement of memory in which fast, small storage is backed by slower, larger storage. Message bus: a shared channel through which agents post and receive messages. Metadata: extra labeled information about a chunk. Meta-harness: a system that combines and controls many agents or many harnesses at once. Metric: a number that measures some aspect of performance, such as speed or cost. Minimum viable tool set: the smallest set of tools that does the job. MLOps: the short form of Machine Learning Operations (MLOps). Mock model: a scripted stand-in for a real model, used so tests behave the same every time. Mock or fake: a controlled stand-in for a real outside component. Model: the trained program, such as a language model, that does the raw reasoning. Model adapter: the harness component that translates between the harness and a particular model provider's format. Model Context Protocol (MCP): an open standard for connecting models to tools and data in a common way. Model routing: sending easy tasks to cheaper models and hard tasks to stronger ones. Model-driven control: control in which the model itself decides what step to take next. Model-graded scoring: grading answers by asking another model to judge them. Monitoring: continuously watching a running system to notice problems. Multi-agent system: a setup in which more than one agent works together on a job.