Reference
Agent-building glossary
The core vocabulary for building real AI agents — in plain English.
- Agent
- An LLM given a goal, tools, and a loop: it reasons, calls tools, observes the results, and repeats until the task is done.
- Tool
- A function an agent can call — search, a calculator, an API. The model emits a structured call; your code runs it and returns the result.
- Tool call
- The model's structured request to invoke a tool, with arguments — the bridge between reasoning and real action.
- Context window
- The finite span of tokens a model can attend to at once; everything the agent “knows” in a turn must fit inside it.
- Memory
- State an agent carries across turns or sessions — facts, preferences, history — beyond what fits in the context window.
- RAG
- Retrieval-Augmented Generation: fetching relevant documents (usually by embedding similarity) and adding them to the prompt so the model answers from real, current data.
- Embedding
- A vector representation of text where similar meanings sit close together, enabling semantic search.
- Vector store
- A database of embeddings that returns the nearest matches to a query vector — the retrieval half of RAG.
- Planning
- Decomposing a goal into ordered steps before or while acting, so an agent handles complex tasks reliably instead of improvising.
- Multi-agent
- Multiple specialized agents that hand off or collaborate — e.g. a planner, a coder, and a reviewer.
- Eval
- A repeatable test of agent quality on held-out cases — the difference between “seems to work” and “works.”
- LLM-as-judge
- Using a model to score another model's output against a rubric — for evals at scale where exact-match comparison won't do.
- Guardrails
- Input and output checks that constrain what an agent accepts or emits — validation, filtering, and refusals.
- Prompt injection
- An attack where untrusted input smuggles in instructions that hijack the agent — the top safety risk for tool-using agents.
- Human-in-the-loop
- Requiring human approval before an agent takes a consequential action, like a payment or a delete.
- Hallucination
- A confident but fabricated answer; mitigated with retrieval, tool use, and evals.
- System prompt
- The standing instructions that define an agent's role, tools, and constraints on every turn.
- Token
- The unit a model reads and writes (roughly ¾ of a word); cost and context limits are both measured in tokens.
- MCP
- Model Context Protocol — an open standard for exposing tools and resources to agents so any client can use them.
- Temperature
- A sampling control: low values make output focused and deterministic, high values make it more varied and creative.