Beyond Prompts & Models: Evolving Enterprise Agentic Systems

As organizations experiment with enterprise agentic systems, discussions often center on models, frameworks, and tools. Far less attention is given to the recurring architectural decisions and engineering trade-offs that shape how these systems evolve over time.

In our session we present a practical architectural evolution model that emerged retrospectively from recurring architectural decisions and trade-offs, observed while iteratively prototyping and evolving enterprise agentic systems. The model organizes these recurring patterns into a series of evolution stages that help engineering teams understand where their architecture is today and which engineering challenges typically emerge next.

Drawing on enterprise AI initiatives at Swisscom spanning product management, software development, and AI-assisted operations, our session traces the architectural decisions that shaped each stage of this evolution: from establishing data readiness as the foundation for context engineering; through rapid low-code prototyping for workflow validation on n8n; to rebuilding multi-agent architectures on the Strands Agents SDK under enterprise deployment constraints; from an early investment in vector-based RAG to a leaner vectorless RAG approach that unblocked us; standardizing agent capabilities via MCP; and finally toward architectural simplification by re-realizing the same workflows with Claude Code, GitHub Copilot and reusable agent skills - thereby trading custom orchestration for a smaller footprint.

Based on real-world implementations, this session demonstrates how successful evolution of enterprise agentic systems is driven less by model capability and more by architectural choices around orchestration, context engineering, interoperability, and architectural simplicity.