Agentic AI

A sparring partner, not an autopilot

We build AI agents. We started in our own marketing, and out of that grew a consulting offer. What we learned sits in an approach of six phases that transfers to your own topics.

Agentic AI Services

AI is not magic. AI is a stack.

Agentic AI is the youngest layer of a long development that begins with rule-based systems and runs through machine learning to generative models. The difference to the layer below: generative AI answers, agentic AI acts.

AI agents consist of few parts. A language model with a system prompt, reusable skills, a knowledge base and the feedback of its users.

Our approach in six phases

What the agent is for – and what it is not

At the start there is no technology, there is a decision. Which role does the agent take, which tasks does it handle and which explicitly not. Without that clarity it stays generic. And what it is there for also determines how critical it is.

To the article:

An agent profile with role and limits – before any technology exists.

The guard rails of the agent

The system prompt describes context, role, tone, way of working and quality standards. It does not make the agent creative, it makes it dependable. This is also where the rules move in that nobody wants to argue about later. Who checks, who approves, where things escalate

An agent that behaves tomorrow the way it does today.

Knowledge beats prompt length

An agent is only as good as the knowledge it can reach. What has proven itself: subject documents, style and format guidelines, planning material. One rule matters most of all. Every fact sits in exactly one file, and all the others point to it.

To the article:

Answers from your own knowledge – and you can see where they come from.

Recurring tasks, consistent results

Skills are predefined working instructions for tasks that come up regularly. Built once, improved iteratively, reused permanently. That produces comparable results without instructing the agent from scratch every time. The knowledge base supplies the what and the why, the skill the how.

It does not write for you. It thinks with you.

In everyday use quality comes out of dialogue. Asking back, disagreeing, sharpening together. One point we underestimated: the better the agent gets, the more conversations it draws away from the team. That can be managed; it cannot be ignored.

To the article:

Better results through follow-up questions—and a team that stays in touch.

Use does not develop an agent. Care does.

Every wrong output is a clue. Usually the fault is not in the model but in a gap in the knowledge base. Spot it, find the cause, sharpen it. And record the result instead of losing it again in the next conversation.

To the article:

An agent that can do more over time – without being rebuilt.

Before you start

Delimitation

AI agents take expert load off your team. They make it possible to work soundly even in areas where nobody on the team is a specialist. That is where we begin. As they mature they take on more, at your pace. Approval stays with people.

Requirements