APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH B Backoff: a retry strategy that waits a little before trying again, often waiting longer after each failure. Baseline: the starting measurement that you compare later changes against to see whether things got better. Batch execution: running many similar jobs together as a group instead of one at a time by hand. Before and after comparison: a test of one change on the same tasks. Benchmark: a shared public test that lets people compare different models or systems on the same set of tasks. Bias: a systematic unfair tilt in outcomes. Bias in AI: a pattern in which a system treats people or groups unfairly because of skewed data or design. Bias in scoring: a tendency of a grader, human or model, to favor answers for reasons unrelated to quality. Blackboard pattern: a way of organizing several agents so they all read from and write to one shared workspace and react to changes there. Blameless postmortem: a review held after a failure that studies what happened and how to prevent it without blaming any person. Blast radius: how much damage one failure can cause. Bottleneck: the step that limits overall speed. Branching: the path splitting depending on a condition. Browser agent: an agent that reads and acts on websites through a browser. Budget: a limit on how much time, money, or text an agent is allowed to use. Budget tracker: the harness component that watches how much of the budget has been used and stops the agent when the limit is reached. Build-versus-buy decision: the choice between making a piece of software yourself and paying someone else for a ready-made one. Business intelligence: using data to understand and guide a business.