The AI agent glossary.

Software that uses a large language model to plan and take multi-step actions toward a goal, calling tools, reading data, and writing back to your systems, escalating to a human when needed.

Unlike a chatbot, an agent decides what to do next and acts, rather than just answering in one turn.

A category of AI systems that pursue goals over multiple steps, choosing actions and tools autonomously rather than responding to a single prompt.

It is the shift from "answer my question" to "get this task done", and it is why integration and guardrails now matter more than the model.

A model trained on large text corpora that predicts and generates language, the reasoning engine inside most AI agents.

The model is rarely the constraint anymore; getting it to act reliably in your systems is.

The mechanism that lets an agent call external functions and APIs, look up an order, update a record, send a message, instead of only producing text.

Tool use is what turns a language model into something that can actually do work.

A pattern where the agent retrieves relevant documents from your knowledge base and grounds its answer in them, with citations.

RAG is how you get answers grounded in your data instead of the model inventing a policy.

An open standard for connecting AI agents to tools, data sources, and systems through a common interface.

MCP makes integration tractable, which is a big reason agents became practical to deploy.

An automated test that measures whether an agent behaves correctly on representative cases, run before any change reaches users.

Evals are to agents what unit tests are to software: no eval, no way to trust a release.

A hard constraint on what an agent may do or say, on inputs, outputs, and risky actions like payments or submissions.

Guardrails plus human approval gates are what make an agent safe to put in front of real users.

Coordinating multiple specialist agents, often under a supervisor, so a complex job is decomposed and completed across steps.

Useful when one workflow needs different skills, retrieval, action, review, handed between agents.

A design where a person approves or reviews the steps that carry real risk, while the agent handles the routine path.

It keeps authority over consequential actions with your team, where it belongs.

The gap between an agent that demos well and one that runs reliably in production, the integration, evaluation, guardrails, and ownership most pilots never cross.

Closing this gap is the hard, valuable part, and the reason many agent projects stall.

Running an agent inside your own virtual private cloud so data, auth, and security stay within your boundary.

It is how regulated and sensitive workloads get an agent without sending data to a vendor.