Agentic Engineering in practice: How AI agents at DATEV are accelerating software modernisation

Agentic Engineering is currently receiving a great deal of attention. How would you describe the concept, and why should organisations focus on it now?

Agentic Engineering refers to the disciplined engineering of software using AI agents. The approach is deliberately different from the purely experimental use of generative AI, where outcomes are often created through isolated prompts and can be difficult to trace or validate.

At its core, the concept addresses a key question: How can organisations leverage the speed of AI without compromising quality, security, or architectural principles? To achieve this, structured practices such as Specification-Driven Development, frameworks like BMAD, and concepts such as Harness Engineering are used. These provide the foundation AI agents need to work not only quickly, but also in a transparent, verifiable, and reliable manner.

The topic is particularly relevant today because many organisations are facing extensive modernisation initiatives while simultaneously dealing with significant time constraints and a shortage of skilled professionals.

What challenge did you and your team face?

As part of its ModFac modernization program, our customer faced the challenge of modernising existing legacy systems significantly faster than would have been possible using traditional development approaches.

This included, among other things, rebuilding existing on-premises components. It quickly became clear that simply introducing AI into the process would not be enough. Architectural requirements, compliance obligations, and quality standards still need to be adhered to consistently.

Agentic Engineering proved to be the right approach. Through clearly defined specifications and a controlled development framework, AI agents could be deployed productively without sacrificing transparency or control.

How was Agentic Engineering applied in the project?

We worked with the BMAD framework, which enables collaboration among multiple specialized AI roles. Agents took responsibility for architecture, implementation, and quality assurance tasks, all based on a shared specification rather than isolated prompts.

The agents generated code, verified requirements against architectural standards, and validated outcomes through automated testing procedures. In addition, a cross-model review process was introduced, where a second AI model independently evaluated the results.

This created a multi-stage process built not on trust, but on verifiable evidence.

What results were achieved?

From a technical perspective, we were able to ensure that generated solutions demonstrably aligned with architectural, security, and quality requirements. Automated validation mechanisms made risks and unresolved decisions visible at an early stage.

Another major benefit was increased transparency. Critical architectural topics, such as storage strategies and integration concepts, could be identified early and addressed consciously.

From a business standpoint, the approach significantly accelerated modernisation efforts. At the same time, business stakeholders, architects, and project leaders retained full control over requirements, decisions, and approvals. Management received a reliable basis for decision-making rather than being asked to rely solely on AI-generated code.

Which challenges are particularly relevant for other organisations?

The most important lesson is this: AI agents are only as effective as the framework in which they operate. Even a high-quality specification is not enough on its own. If documentation is incomplete, architectural principles are unclear, or control mechanisms are missing, agents may make incorrect assumptions or prematurely consider tasks complete.

This becomes especially apparent in complex modernization programs, where a well-defined harness is essential. Organizations should therefore invest sufficient time in preparation and ensure that governance, quality assurance, and documentation are considered from the outset.

What prerequisites should organisations have before getting started?

The most important success factor is clarity. Before deploying an AI agent, organisations must clearly define the target state, expected value, success criteria, and operational boundaries.

Robust project documentation, architectural guidelines, automated testing, and appropriate access controls are equally important.

Most importantly, the human role remains critical. AI can support and accelerate development, but responsibility for decisions and approvals must remain with experienced engineering and architecture teams.

What use cases are particularly well suited for Agentic Engineering?

We see especially strong potential in modernisation and transformation initiatives. These projects often involve replacing highly complex legacy systems that have evolved over many years and contain extensive business logic.

The approach also offers significant advantages in regulated industries such as financial services, insurance, and public administration. The combination of specification, governance, and automated validation provides the traceability required to use AI productively and responsibly.

In general, Agentic Engineering is a strong fit wherever quality, documentation, and auditability are just as important as speed.

How do you see this field evolving over the coming years?

I believe Agentic Engineering will follow a trajectory similar to DevOps or agile methodologies. What is still often viewed as a pilot initiative today will increasingly become the industry standard.

Frameworks such as BMAD and concepts like Harness Engineering will continue to mature, become more standardised, and integrate more easily into existing development environments.

At the same time, the importance of software architects will continue to grow. Their role will increasingly be to define the guardrails within which AI agents can operate safely, efficiently, and responsibly.