Agentic Engineering: Building Sovereign Software with B-MAT

The Paradigm Shift: From Vibe-Coding to Agentic Engineering

The conversation highlights a critical evolution in software development: the transition from 'vibe-coding'—where users prompt AI for immediate, often unpolished results—to 'agentic engineering.' While vibe-coding offers speed, it lacks the architectural integrity, quality controls, and testability required for enterprise-grade applications. Agentic engineering, by contrast, utilizes AI agents as structured subordinates within a rigorous development framework. This approach allows a single architect to oversee the production of complex systems that previously required large engineering teams.

The B-MAT Framework: Structured AI Development

Central to this shift is the B-MAT framework, which formalizes the interaction between human architects and AI agents. Instead of letting the AI decide the architecture, the human provides a Product Description (PAD) and a clear architectural blueprint. The framework then utilizes 'skills'—pre-defined contexts, templates, and standards—to guide the agents. This ensures the AI acts as an interviewer and executor rather than an autonomous decision-maker. By maintaining the human as the Product Owner, the B-MAT framework ensures that the final output aligns with specific business goals and quality standards.

Strategic Digital Sovereignty and the NEMU Case Study

A primary driver for this methodology is 'digital sovereignty'—the strategic move to reduce dependency on third-party 'commodity' tools like Slack, Teams, or Zoom. The NEMU project serves as a practical case study: a sovereign chat solution built on the Matrix protocol. By using agentic engineering, the developer was able to build a solution that includes enterprise-critical features—such as OpenID Connect, compliance flows, and scalability—without the vendor lock-in or high licensing costs of commercial alternatives. This 'Make' strategy is now viable because AI can bridge the gap between a 90% complete open-source project and a 100% compliant enterprise tool.

Architectural Optimization for AI

A key technical insight shared is the need to optimize code for the LLM rather than for human readability. Traditional software often uses complex inheritance and heavy frameworks to manage human cognitive limits. However, these structures can confuse AI agents. By designing 'flat' architectures—where the logic is simplified and the complexity is handled in SQL scripts or explicit specifications—the developer can produce more reliable and maintainable code. In this model, the human focuses on the 'test' as the ultimate scripted specification. If the test passes, the underlying code is considered successful, shifting the focus from manual code review to automated verification.

Conclusion: The Human as the Pilot

Ultimately, the role of the human in the age of agentic engineering is that of the 'pilot.' While the AI handles the heavy lifting of code generation and technical implementation, the human remains responsible for the vision, the architectural constraints, and the final quality assurance. By investing in high-quality specifications and agentic skills, organizations can achieve a level of independence and operational agility that was previously unattainable for small teams.