Appendix A: Glossary

Term	Definition
Agent	A system that repeatedly uses an LLM, acts through tools, updates state, and continues until a stop condition is met
Artifact	A produced object such as a file, report, image, query result, or generated dataset
Context window	The finite ordered set of text units provided to the model for one call
Token	The basic unit of text a model processes. Tokens are typically 1–4 characters long and can be a word, part of a word, or punctuation.
Tokenization	The process of splitting text into tokens before the model processes it.
Final answer	The agent’s declared completion of the task
Guardrail	A deterministic control that validates, constrains, or blocks agent behavior
Harness	The deterministic, stateful runtime that wraps the LLM engine and turns it into an agent
LLM engine	The stateless model service that predicts text from context
Observation	The result of a tool action, returned to the agent’s context
Orchestration	Coordination of multiple agents or subtasks
ReAct	A loop pattern that alternates reasoning, action, and observation
Span	A timed unit of work inside an agent trace
State key	A reference to stored state or an artifact, used instead of pasting large content into context
Structured output	Model output that follows a parseable contract
Tool	A controlled capability given to the agent with an explicit contract
Trace	A hierarchical record of an agent run and its steps