Agentic engineering

Building an agent takes an afternoon. Building an agent that still works reliably a year later, that somebody owns and whose permissions the organization understands, is a craft. We call that craft agentic engineering: the discipline of designing, building, testing and managing AI agents so they keep working inside an organization.

The need is concrete. Many organizations now have dozens of agents, built by enthusiastic staff, with no overview of what they do or which data they touch. Others have a single agent in production that is allowed too much, whose outputs nobody checks and whose costs surprise them every month. In both cases the same layer is missing: a way of working for designing and managing agents.

What you learn

The training follows the life cycle of an agent. Day one is about design. What sets an agent apart from a chatbot or a script, and when a plain script is the better choice. How to break a task into steps, which tools an agent needs and how to build them so the agent can call them safely. How to manage context and memory. And the question that returns in every design: where does the human decide, and how do you enforce that technically instead of leaving it to an instruction.

Day two covers building, testing and managing. Participants build an agent on the platform their organization uses and learn to evaluate it systematically: creating test sets, measuring quality, preventing regressions when the model or the instructions change. We cover the risks specific to agents, such as prompt injection through documents and e-mails, unwanted access to data and actions outside the intended scope. Then comes the production phase: logging, monitoring, incidents, version control and the moment you switch an agent off.

Agents in the organization

Technology is half the job. An agent that goes into production belongs in the organization's AI register, with an owner, a purpose, a risk class and an approval. For applications in a high risk class, the AI Act requires human oversight, documentation and logging. In the final half-day we translate this into an agent roadmap for your own organization: which agents first, which agreements go with them and what it takes to sustain the maintenance. The approach matches how we set up AI implementation and AI governance.

Format

The training is vendor-independent. We work with Copilot Studio, Azure AI Foundry or an open framework, depending on what your organization uses. Participants bring their own case: a process in which an agent can add value. Over two days they take that case from design to a tested first version with a maintenance plan. For an in-company training we agree the cases and the platform in advance.

Prior knowledge

Participants have experience building a first agent, for example from Introduction to Copilot Studio, or a comparable technical background. Programming experience is useful and not required; the emphasis is on design choices, evaluation and maintenance.