From SOA to Agents: What Three Turns of Platform Extensibility Teach Architects

Abstract

Enterprise platforms have gone through three distinct extensibility models over the last two decades: service-oriented architectures built on SOAP and WSDL, RESTful microservices built on OpenAPI contracts, and, most recently, agentic systems built on natural-language intent and tool-invocation schemas. Each transition has been described in its own moment as an unprecedented break from what came before. This paper argues that framing is only partly correct. Drawing on a comparative qualitative analysis across all three eras, informed by first-hand architectural experience with SOAP/WCF integration, Azure-based microservice platforms, and an ongoing agentic extensibility initiative on the Microsoft Azure Portal, this paper identifies four principles that have remained constant across all three transitions: the interface a platform exposes is the product regardless of its notation; how a platform evolves that interface without breaking existing consumers determines the extensibility model's long-term viability; trust and permission boundaries must be treated as architecture rather than an added layer; and partner-facing ergonomics, not raw platform capability, decide whether an extensibility model is actually adopted. The analysis also identifies what is genuinely new in the agentic era: non-deterministic outputs that cannot be verified through conventional unit testing, capability contracts that must encode intent and reversibility rather than only parameters and return types, and trust boundaries that must now govern what an agent is permitted to decide, not only what data it can access. The paper concludes with a practical checklist for architects evaluating agentic platforms, and argues that engineers who have already navigated one or two of these transitions carry transferable judgment into the third, provided they distinguish genuinely new problems from old problems wearing new vocabulary.