Agentic AI, in plain terms

What is agentic AI?

Also called: agentic artificial intelligence, AI agents

Agentic AI is software in which a large language model does not just produce text but decides what to do next: it plans steps toward a goal, calls tools such as search, databases or APIs, observes the results, and continues until the task is done or it needs a person.

The difference from a chatbot is action. A chatbot answers a question; an agentic system can look up the order, draft the reply, update the record and ask for approval before issuing the refund. Because it acts, it needs controls a chatbot does not: permissions, spending limits, guardrails and a human in the loop for consequential steps.

What is an AI agent?

Also called: LLM agent, autonomous agent

An AI agent is a program that uses a large language model to choose and carry out actions toward a goal. It combines instructions that define its role, a set of tools it is allowed to call, memory or knowledge it can consult, and a loop that alternates between reasoning and acting.

The reason-then-act loop was formalised in the ReAct paper: the model writes a thought, takes an action, reads the observation, and repeats. Most production agents are a variation on that loop with stricter limits on which tools may run and when.

What is the Model Context Protocol (MCP)?

Also called: MCP, MCP server, MCP client

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in November 2024, for connecting AI applications to external tools and data. An MCP server exposes tools, resources and prompts in a standard format; an MCP client, such as an AI assistant or an agent platform, discovers and calls them without custom integration code for each one.

MCP is often described as a USB-C port for AI: one connector shape instead of a bespoke integration per tool. Servers can run locally over standard input/output or remotely over HTTP.

What is retrieval-augmented generation (RAG)?

Also called: RAG, grounded generation, knowledge base search

Retrieval-augmented generation (RAG) is a technique in which a language model is given relevant passages retrieved from an external document collection at the moment it answers, so the response is grounded in those documents rather than only in what the model memorised during training.

A typical pipeline splits documents into chunks, converts each chunk to a numeric embedding, finds the chunks closest to the question, optionally re-ranks them, and passes the best ones to the model along with the question. RAG reduces fabricated answers and lets a model use private or recent information without retraining.

What are AI guardrails?

Also called: LLM guardrails, content safety filters

AI guardrails are controls that sit outside a language model and check what goes into it and what comes out of it — for example detecting personal data, prompt-injection and jailbreak attempts, toxic language or policy violations — and then block, redact or escalate. They are enforced by the surrounding system rather than requested in the prompt.

What is human-in-the-loop approval?

Also called: HITL, human oversight, approval gate

Human-in-the-loop approval means an automated system pauses before a consequential action — sending money, deleting data, contacting a customer — and waits for a person to approve or reject it. It keeps a human accountable for decisions the system is not trusted to make alone.

Human oversight is a legal requirement for high-risk AI systems under Article 14 of the EU AI Act, and a recommended control in most AI risk frameworks.

What is AI workflow automation?

Also called: agentic workflow, AI workflow

AI workflow automation is the design of a multi-step business process as an explicit graph of steps — triggers, conditions, tool calls, approvals — in which some steps are carried out by a language model or an AI agent. The graph fixes the order and the control points; the model handles the steps that need judgement or language.

A workflow is the more predictable of the two common patterns. In a workflow the path is drawn in advance; in a free-running agent the model picks the path. Many production systems combine them: a workflow for the process, agents for individual steps.

What is the Agent-to-Agent (A2A) protocol?

Also called: A2A, Agent2Agent, agent card

The Agent-to-Agent (A2A) protocol is an open standard that lets AI agents built by different vendors discover each other and exchange tasks. Each agent publishes an agent card — a JSON document, conventionally at /.well-known/agent-card.json — describing what it can do and how to call it.

A2A and MCP are complementary: MCP connects an agent to tools and data, while A2A connects an agent to other agents.

What is AG-UI?

Also called: AG-UI protocol

AG-UI is an open, event-based protocol for the connection between an AI agent and the user interface in front of it. It standardises how an agent streams messages, tool calls and state changes to a front end, so any compliant interface can drive any compliant agent.

What does bring your own key (BYOK) mean?

Also called: BYOK, bring your own model, model-agnostic

Bring your own key (BYOK) is an arrangement in which a customer supplies their own API credentials for a language-model provider to a platform that runs on top of it. The customer keeps the billing relationship and the choice of model, and the platform is not a reseller of model usage.

What is multi-tenancy, and what is schema-per-tenant?

Also called: tenant isolation, schema per tenant

Multi-tenancy is an architecture in which one running application serves many customers (tenants) while keeping each tenant's data separate. In the schema-per-tenant model every tenant gets its own database schema — its own set of tables — inside a shared database, which isolates data more strongly than a shared table filtered by a tenant column.

What is LLM-as-a-judge evaluation?

Also called: LLM judge, model-graded evaluation

LLM-as-a-judge is an evaluation method in which one language model scores the output of another against stated criteria such as accuracy, relevance or tone. It makes it practical to grade large volumes of open-ended answers, at the cost of inheriting the judging model's own biases.