AI Agent & LLM Glossary

The complete set of actions an agent is authorized to take, defined by its registered tools and any policy filters.

The wrapper layer around an LLM that adds the agentic loop, tool routing, memory, and observability.

The perceive, reason, act, observe cycle that an autonomous agent runs until the goal is reached or a stop condition fires.

A system that uses an LLM to decide actions, call tools, and iterate toward a goal. Not just generate text.

The component that carries out planned actions by invoking tools, calling APIs, and reporting results back.

A workflow where a human reviews or approves agent decisions inline, before the agent proceeds to the next step.

A workflow where humans supervise an autonomous agent and intervene only when monitoring thresholds are exceeded.

Two or more specialized agents that collaborate, hand off tasks, or critique each other.

The top-level agent or controller that routes work to specialists in a multi-agent system.

The component that decomposes a high-level goal into an ordered sequence of executable sub-tasks.

Reason → Act → Observe loop where the LLM verbalizes its reasoning before each tool call.

A specialist agent invoked by an orchestrator with a narrow tool set and focused prompt.

A multi-agent topology where one supervisor agent routes incoming tasks to specialist worker agents and aggregates their results.

The maximum number of tokens (input + output) a model can process in a single request.

Persistent storage outside the context window. Typically a vector database, key-value store, or both.

Reusing previously-processed input tokens at a 90 percent cost discount and lower latency.

The conversation history kept in the LLM's context window during a single session.

Open standard from Google for communication between independent AI agents from different vendors.

IBM-originated open standard for AI agent communication, now merged into A2A under the Linux Foundation.

JSON metadata document advertising an A2A agent's skills, endpoint, and authentication requirements.

OpenAI's name for tool use. Same concept, slightly different API surface.

An open standard from Anthropic for two-way connections between AI applications and external data sources, tools, and prompts.

A connector inside the Host that maintains a 1:1 stateful session with exactly one MCP Server.

The AI application that runs the LLM and creates one or more MCP Clients to talk to external Servers.

A program that exposes Tools, Resources, and Prompts to MCP Clients over JSON-RPC.

The LLM's ability to call your code by emitting structured function calls instead of free-text answers.

Splitting documents into smaller pieces before embedding so retrieval returns relevant fragments.

A vector representation of text that captures semantic meaning. Similar text gets similar vectors.

Fetching relevant documents at query time and injecting them into the LLM prompt to ground answers in real data.

A second-stage model that re-orders retrieved chunks by true relevance, not just embedding similarity.

Finding documents by meaning, not by keyword overlap, using embedding similarity.

A database optimized for storing embeddings and finding nearest neighbors at scale.

How a voice agent responds when the caller starts speaking while the agent is still talking (a barge-in event).

Legacy phone menu technology where callers press digits or speak short keywords to navigate predefined branches.

The allocation of acceptable delay across each stage of a voice agent pipeline so the total Time to First Audio meets the target.

The deep learning model inside a TTS system that converts predicted acoustic features into actual audio waveforms.

The duration of detected silence (typically 300 to 800ms) that triggers the agent to consider the user's turn complete.

The first stage of a voice agent pipeline that transcribes user audio into text in real time using deep neural networks.

XML-based markup that controls how a TTS engine pronounces text, including pacing, emphasis, pronunciation, and pauses.

The three-stage architecture of every modern voice agent: speech to text, then language model, then text to speech.

Neural model that synthesizes natural-sounding speech audio from the LLM's text output, streamed back to the caller.

The total latency from when the user stops speaking to when the agent's first audio chunk plays back. The single most important voice agent metric.

How a voice agent decides when the caller has stopped speaking and it is the agent's turn to respond.

A voice bot reads scripted responses on detected keywords. A voice agent reasons, plans, and takes actions across multiple tools.

Using a few seconds of reference audio to synthesize new speech in that specific voice via TTS.

Continuing training on your own data to adjust the base model's behavior or knowledge.

A neural network trained on text that predicts the next token, used as the engine inside agents.

Training method that uses human preference rankings to align model outputs with human values.

A test suite that runs the model against a fixed set of inputs and grades outputs automatically.

Tying model outputs to verifiable sources. Usually retrieved documents or tool results.

When an LLM generates plausible-sounding text that is factually wrong or invented.

Input and output filters that prevent unsafe, off-topic, or out-of-policy model behavior.

An attack where user input contains instructions that hijack the LLM's behavior.

Marketing legacy software, simple chatbots, or basic automation as "AI agents" without the autonomous decision-making that real agents have.

The disciplined production successor to vibe coding: AI agents propose changes, humans review, tests gate merges, decisions get documented.

The discipline of designing what an LLM sees at inference time, including system prompt, retrieved data, tool definitions, and conversation history.

Retry strategy that doubles the wait between attempts to give a struggling system room to recover.

Logging, metrics, and tracing for LLM calls so you can debug, audit, and optimize cost.

When an AI initiative stays stuck indefinitely in the pilot or proof-of-concept phase and never reaches production deployment.

Capping how many requests a user, IP, or app can make in a time window.

Andrej Karpathy's term for building software by describing what you want to an AI and iterating on the result without reading every line.