AI Basics: Key Terms Explained

AI agents are more than chatbots. They combine language models with instructions, context, tools, and permissions to get things done in your work environment. This glossary explains the key terms and shows how they work together.

---

LLM: The Language Model Behind AI

LLM stands for "Large Language Model". An LLM is an AI model trained on massive amounts of text. You give it a prompt and it predicts the next word step by step until the answer is complete.

Well-known models come from OpenAI, Anthropic, or Google, for example. Provider and model are not the same thing. A provider can offer multiple models with different strengths, speeds, and costs.

Where a model runs doesn't depend solely on the provider. The same model can run in different regions or data centers depending on the provider, for example in the EU or in Germany. This means models can be used in compliance with GDPR independent of their original provider, under certain conditions.

An LLM by itself can generate text. It can't use tools, take action, or access your project data on its own. That's where the other building blocks come in.

Prompt

A prompt is the input you give to the LLM to assign it a task. It can include a question, an instruction, or additional context, for example. The LLM processes this text and generates its response.

Reasoning

Reasoning refers to an AI model's ability to handle complex tasks across multiple processing steps. The model can analyze information, make connections, and develop a solution step by step.

Depending on the model and configuration, reasoning can be more or less pronounced. This can affect the quality, speed, and cost of a response.

Stream

Stream refers to the step-by-step delivery of an AI response as it's being generated. Instead of waiting for the complete response, parts that have already been generated are displayed right away.

This means you see initial results sooner, even while the full processing is still running.

---

Chatbot: Ask Questions and Get Answers

A chatbot is an LLM with a user interface. You type something, and the chatbot answers in natural language.

In addition to the LLM, a chatbot can have a system prompt and memory. This lets it take on a fixed role or remember information across multiple messages.

The key difference from an agent: A chatbot responds, but it doesn't act independently. For example, it won't create a task on its own or send an email.

Chatbots work best for tasks where you provide the necessary context directly:

• Quick text work like rephrasing or creating variations

• Research for a specific question

• Summaries of content you provide yourself

• Brainstorming and idea checks

---

AI Agent: AI That Gets Things Done

An AI agent is a configurable digital assistant. It consists of a language model equipped with an identity, instructions, access to selected information, and optional capabilities. This lets it plan tasks, execute multiple steps, and work with its environment.

The difference is easy to remember:

• Chatbot: You ask, it answers

• Agent: You set a goal, it works toward it

An agent can decide on its own which information or tools it needs for the next step. It works in a loop: plan, act, check the result, and continue until the goal is reached. This so-called agent loop runs in a protected environment called a sandbox.

What building blocks make up an agent?

An agent consists of several building blocks that work together:

• LLM For processing and text generation

• System Prompt For role, rules, and objectives

• Context For the knowledge the agent needs for the task

• Tools and MCP So the agent can act and use external systems

• Skills For repeatable and consistent approaches

• Memory For stored information over time

• Trigger So the agent starts on demand, on schedule, or when an event occurs

• Human in the Loop For defined human approvals at critical points

Important: More autonomy can make an agent more useful. At the same time, the demands on permissions, boundaries, control, and transparency increase.

---

System Prompt: The Fixed Guidelines

The system prompt is a fixed instruction in the background. It sets what role an agent has, how it should work, and what goal it pursues.

This turns a general language model into a specialized agent. A system prompt can include specifications for role, tone, guidelines, and working methods.

---

Context: The Knowledge for the Current Task

Context is the sum of all information available to the LLM for a task. This can include your prompt, files, the conversation history so far, or connected tools.

More context isn't automatically better. Information that's relevant and carefully prepared gives better results than a large amount of contradictory or irrelevant data. Also, each model has a limited context window.

---

Memory: Stored Context

Memory refers to stored context connected with an agent. It can contain information or previous interactions that get reused in future conversations.

The difference from normal context: Memory persists across individual tasks or conversations. Not every piece of information needs to be stored permanently.

---

Skill: Reusable Instruction

A skill is a reusable instruction. It describes how a task should be completed according to a specific standard.

For example, an LLM can write code. A skill can additionally define what structure, naming conventions, and tests apply. For an agency, a skill might define how a brief is reviewed according to their own template.

Skills help with three things in particular:

• Consistency The task gets completed according to your desired standard

• Reliability and efficiency Repeated fine-tuning and explanations are reduced

• Reuse and sharing A skill created once can be used by your team

• Company-specific customization Your own templates, processes, terminology

• Modular and expandable New capabilities can be added without changing the core system

Skills extend an agent's capabilities without you having to change its basic configuration.

Skills vs. Tools

• Tool: What an agent can do, like a web search

• Skill: The instruction for how to use that capability correctly

---

Tools and Capabilities: What an Agent Can Use

Capabilities is the umbrella term for the abilities available to an agent. This includes skills, built-in tools, and connectors.

The core principle is: Capabilities are explicitly granted, not simply assumed. You decide which tools, context, and permissions an agent receives. This gives you control over what the agent can and can't do.

---

Connectors and Connections: Links to Other Systems

A connector establishes the connection to external services or systems. This lets an agent access data or functions outside of awork.

A connection is a single authenticated identity for that connector.

The connector's connection status and whether it's linked to a specific agent are managed separately.

---

MCP: The Shared Interface for AI

MCP stands for "Model Context Protocol". It's an open standard that connects AI agents to external tools and data sources through a shared interface.

Simply put, MCP is like USB-C for AI: a standard that many systems can plug into. Technically, MCP links a tool's API with the agent. This lets the agent read data or trigger actions without needing a custom integration built for each tool.

MCP is especially relevant when an agent needs to work across multiple systems, like with a project management tool, CRM, or Slack.

---

Trigger and Schedules: What Starts an Agent

A trigger determines when an agent becomes active. It can start on demand or be triggered by an event or schedule.

A schedule is for recurring or proactive tasks. An agent can be configured to perform certain work regularly without you having to manually start it each time.

For example, an agent could react to a new file or regularly run a specific workflow.

---

Human in the Loop: Human Control at Critical Points

Human in the Loop is a deliberately placed control point. The agent stops and waits for a human decision before continuing.

You set in advance where this approval is needed. For example, an agent can set a status or prepare a document, while an email to a customer is only sent after human approval.

The principle is simple: The agent works, the human decides at the critical points.

---

Thread and Run: How Agent Executions Are Documented

A thread is a conversation history with an agent. Threads document the exchange between you and the agent and let you trace previous conversations. A scheduled agent run also appears as a thread.

A run is a single execution within a thread. A thread can contain multiple runs. Each run has an owner.

---

How the AI Building Blocks Work Together

An agent works in a loop.

1. A trigger starts the agent

2. The LLM processes the system prompt, context, and task

3. The agent uses tools or skills as needed

4. The result feeds back and the agent decides on the next step

5. At defined points, it waits for a Human in the Loop

6. Relevant information can be retained in Memory

This is exactly what sets agents apart from chatbots: The chatbot generates an answer. The agent can work through a task across multiple steps.

---

FAQ

What is a skill?

A skill is a reusable instruction that defines how an agent should complete a task according to a specific standard.

What is MCP?

MCP is an open standard that connects AI agents to external tools and data sources through a shared interface.

Why does an agent need Human in the Loop?

So that a human can make decisions at important points before the agent continues. This lets you intentionally add human approval to critical actions.

What's the difference between a tool and a skill?

A tool gives an agent a capability. A skill describes how the agent should use that capability or complete a task according to a specific standard.

What is a thread?

A thread is the history of an interaction with an agent. Scheduled agent runs are also displayed as threads.

What is a run?

A run is a single execution within a thread. A thread can contain multiple runs.