Specification-Governed Agentic Systems: A Lifecycle and Multi-Plane Reference Framework for Enterprise Multi-Agent Engineering

Enterprises now build many agentic applications on Large Language Models (LLMs), yet a large share of these efforts stall before dependable production, and the cause is rarely raw model capability (Brown et al., 2020; Bommasani et al., 2021). The cause is an engineering one. Agent prototypes tend to couple the parts that rarely change, such as graph topology and module boundaries, with the parts that change often, such as prompts, thresholds and decision policies, so a small behavioural request forces edits across the codebase and erodes the trust that production demands (Parnas, 1972; Sculley et al., 2015). This paper presents a reference framework for building agentic systems as a disciplined engineering practice. The framework rests on three contributions. First, a four-phase lifecycle (define, design, develop, deploy) gives a repeatable path from a framed problem to a running service. Second, a structure-behaviour separation principle splits a frozen scaffold and a canonical specification from a versioned runtime configuration, and pairs a no-code behaviour loop with a gated evidence-to-code loop, so most change is resolved without touching structure. Third, a multi-plane execution model separates an orchestration plane over a typed shared state, a communication plane with a typed message contract, and a tooling plane with a uniform invocation interface, above a governed data-product substrate. A capability and operating-model dimension and a five-level maturity model accompany the framework. A conceptual synthesis was followed by a small expert appraisal, in which practitioners rated change-locality and governance fit favourably (M = 5.8, SD = 0.7, on a seven-point scale). The findings suggest that treating structure and behaviour as distinct change surfaces, under typed orchestration and evidence-driven evaluation, may raise the share of agentic pilots that reach reliable operation, warranting cautious generalisation beyond the appraisal sample.