A prompt is the input text given to a language model to guide what it generates.

A system prompt sets persistent instructions and behavior for a model before the user's messages begin.

A user prompt is the specific request or question a user sends to the model within a conversation.

Zero-shot prompting asks a model to perform a task with instructions alone and no examples.

Few-shot prompting includes a small number of examples in the prompt to show the model the desired pattern.

Chain-of-thought prompting asks a model to reason step by step before giving a final answer.

Prompt templates are reusable prompt structures with variable placeholders filled in at runtime.

Structured output constrains a model to return data in a defined format, such as JSON, instead of free-form text.

Prompt injection is an attack where untrusted text manipulates a model into ignoring its original instructions.

Function calling lets a model request that a specific function be executed with structured arguments it generates.

Tool calling is the general pattern of a model invoking external tools to take actions or fetch information it can't produce on its own.

A tool schema describes a tool's name, purpose, and expected input format so a model knows how to call it correctly.

API tools wrap external APIs so a model can call real services, such as search, weather, or internal systems, as tools.

An AI agent uses a model together with tools, planning, and memory to pursue a goal across multiple steps.

Agentic AI describes systems designed to act autonomously toward goals rather than simply responding to a single prompt.

The agent loop is the repeating cycle of observing state, deciding an action, acting, and evaluating the result.

Planning is how an agent breaks a goal down into an ordered sequence of steps it can execute.

Tool use is an agent's ability to call external tools when it needs information or capabilities beyond the model itself.

Memory lets an agent retain information across steps or sessions instead of treating every interaction as isolated.

Reflection is an agent reviewing its own output or actions to catch errors and improve before continuing.

Human-in-the-loop keeps a person able to review, approve, or override an agent's actions before they take effect.

Multi-agent systems coordinate several specialized agents that collaborate to complete a larger task.

Agent orchestration manages how multiple agents or steps are sequenced, routed, and coordinated within a system.

Frameworks such as LangGraph, CrewAI, and Semantic Kernel provide the plumbing for agent state, tool calls, and control flow.

An agent's state — conversation history, intermediate results, tool outputs — has to be persisted and passed between steps of the agent loop.

Production agent memory is usually backed by a database or vector store, with explicit rules for what gets kept, summarized, or dropped.

Frameworks like CrewAI and LangGraph provide the routing and delegation logic that coordinates several specialized agents.

In production, human-in-the-loop means a real approval queue, notification system, and timeout policy — not just a design principle.

Evaluating an agent means checking whether it chose the right tools and reached a correct outcome, not just whether one response looked reasonable.

Observing an agent means tracing every step of its loop — each decision, tool call, and intermediate result — not just the final answer.

Workflow orchestration sequences AI steps alongside regular application steps — some deterministic, some model-driven.

GenAI introduces new risks beyond traditional application security, because a model treats everything in its context as text it might act on.

A jailbreak is an input crafted specifically to bypass a model’s safety training or stated instructions.

RAG security means treating retrieved documents as untrusted data and filtering retrieval by the requesting user’s actual permissions.

Agent security means scoping tool access tightly and requiring approval for irreversible actions, since an agent turns a decision into a real effect.

Preventing data leakage means detecting and redacting personal information before it reaches a prompt, a log, or a model’s output.

AI services need their own identity and access controls — which service can call which model, and with which permissions.

AI supply chain security means verifying the provenance of third-party models, datasets, and framework dependencies before trusting them.

Red teaming an AI system means deliberately attacking it with adversarial inputs before real attackers do.

AI governance sets organizational policy for what models and data can be used, by whom, and under what oversight.

Responsible AI considers fairness, transparency, and the real-world impact of a system’s outputs, not just whether it works technically.