TL;DR: An AI agent isn't a magic robot – it's a language model with a few building blocks around it: system prompt, context, LLM, tools, skills, memory and a trigger that sets it off. Once you understand these elements, you can decide which tasks are a good fit for an agent.
At the end of this article you'll find a workshop template for building a shared understanding of agents as a team.
[.no-toc]An agent seems super technical, but it doesn't have to be[.no-toc]
The word "agent" sounds like developer territory at first. But at its core, an agent is just a new team member with a few handy building blocks in their pocket. If you can onboard a new team member, you can design an agent.
From chatbot to agent
Almost everyone in agencies has long been using AI in chat – writing copy, researching, summarising.
The next step is agents: AI that doesn't just answer, but takes on tasks independently. And this is exactly where many people hesitate. The word alone sounds highly technical, like something you couldn't shape yourself.
The quickest way to shed that hesitation: think an agent through in the offline world first. A bus driver, a paramedic, a ticket seller – they can all be broken down into exactly the same building blocks as an AI agent. With the help of these analogue metaphors, you'll ease your way in nicely.
[.b-important-block]A chatbot like ChatGPT or Claude waits for your message, replies with text, and then waits again. It tells you how to word an email. It doesn't send it. An agent can act. It decides for itself which steps are needed, reaches for tools, carries out actions and keeps going until the task is done. The difference isn't the model itself, but the building blocks you build around it.[.b-important-block]
[$tag]💡 Reminder[$tag]
What is an AI agent made of?
An agent is made up of eight building blocks: LLM, system prompt, context, tools (plus MCP), skills, memory, trigger and – not a technical building block, but indispensable – the human in the loop.
1. The LLM: the brain
The large language model predicts the next word, one word at a time, and produces text that way. Well-known models come from OpenAI (GPT), Anthropic (Claude) or Google (Gemini). The key question with any agent: fast or smart? Fast, cheap models for simple routines; powerful ones for complex thinking.
2. The system prompt: role and remit
The system prompt is the fixed instruction that tells the model who it is and what its job is – for example: "You are a support assistant for a customer team. Reply briefly and always with a concrete next step." It sets the role, the guardrails and the goal, turning a generic model into a specialised colleague.
3. The context: what the agent currently "knows"
A model only knows what's in the current message. This working memory is called the context window; it's limited, and it's measured in tokens. That's why long chats get "fuzzy". For an agent, the right context matters: agency, client, brand and project knowledge.
4. Tools and MCP: what the agent acts with
Tools are the agent's implements: web search, data access, actions like "send message" or "create entry". Without tools, an agent stays a chatbot. MCP (Model Context Protocol) is a bit like USB-C for AI: a standard for connecting agents to your systems such as CRM, Slack or your project tool. How that plays out together is covered in the article How LLMs, agents & skills work together.
5. Skills: reliable routines
A skill is a packaged instruction for how to do something properly. The difference between "it can write code" and "it knows what good code means for us". Skills make sure results come out consistently to your standard, rather than generic.
6. Memory: what stays over time
Memory is what an agent remembers across individual tasks: guidelines, recurring facts, past decisions. Without memory, every run is a fresh start. The important question: what should actually be remembered permanently? Not everything belongs in long-term memory.
7. The trigger: what starts the agent
The trigger answers: when does the agent start? Either on demand (you give an instruction) or on a schedule or event (every Monday at 9am, or when a new file comes in). The trigger decides whether an agent helps on demand or runs independently in the background.
The important addition: human in the loop
Not a technical building block, but indispensable: the human in the loop. This is where you define at which point the agent has to get sign-off from a human before it carries on. The more capable the agent, the more important this checkpoint – for anything expensive or publicly visible.
How the building blocks work together
The building blocks aren't an assembly line, they're a loop.
- A trigger starts the agent.
- The LLM reads the system prompt, context and task, and plans.
- If it needs a tool or a skill, it calls them, the result flows back, and the model decides: done or keep going?
- At defined points the human in the loop steps in, and anything relevant lands in memory.
This cycle – plan, act, check, repeat – is the real heart of an agent.
Example: the festival bus driver
The whole thing becomes tangible fastest with an example that has nothing to do with AI at first: a bus driver on a festival site.
- Trigger: Drive to the timetable.
- System prompt: "You're a bus driver. Get your passengers safely from A to B."
- LLM: a fast and reliable model – reaction time is what counts here.
- Context: a 15-metre bus in Germany, room for 50 people, the current time and location.
- Tools: driving, seeing traffic, talking to people, reading GPS and a map.
- Skills: reverse-parking a 15-metre bus.
- Memory: the timetable.
- Human in the loop: in an accident, don't drive off – call the emergency services and wait.
Every building block has a clear counterpart. Following exactly this pattern, you can also think through an agent for a real agency task – a briefing assistant, say, or a reporting helper.
Now try it yourself: the workshop template
Got the theory? The best next step is to do it once yourself – ideally together as a team. That's what the "Meet your first AI Agent" workshop template is for: a repeatable format that deliberately takes you through the offline world first, before moving on to your real cases.
👉 To the workshop template "Meet your first AI Agent"
FAQ
Why agents matter for agencies?
80% of agency people use AI on more than three days a week, and only 2% never. But only 16% consider their agency's structures fully AI-ready. That's according to the Agency Happiness Report 2026, which surveyed 500 people from the industry.
The gap is clear: AI has arrived in day-to-day work, but the understanding of what's possible beyond chat is lagging behind. The next step after "using AI in chat" is agents. And to use an agent sensibly – or build one yourself – you need to know what it's made of.
What's the difference between a chatbot and an agent?
A chatbot replies with text and then waits. An agent acts: it uses tools and context, carries out several steps on its own, and works until the task is done.
How do skills and tools differ?
A tool is the ability to do something (e.g. a web search). A skill is the instruction for how to do it properly (e.g. what a good search report looks like for you).
Do I need coding skills to build an agent?
For simple, home-grown agents, increasingly no. What you mainly need is the right platform and a clear understanding of the building blocks. That understanding is the first step – and the workshop template helps with it.
