This paper proposes a technology‑neutral reference architecture for integrating Agentic AI into enterprise software. Agentic AI unifies reasoning, enterprise context, model inference, and external actions into system‑level capabilities. Embedding these capabilities raises architectural concerns across orchestration, knowledge and memory, model access, tool/action libraries, security, runtime infrastructure, and evaluation. Existing academic and industry solutions address many of these dimensions but differ in scope, abstraction level, responsibility organization, interaction representation, and assumptions about required capabilities. Building on this, we synthesize ten responsibilities:
- Orchestration & Scheduling
- Knowledge Management
- Persistent Memory
- Model Access
- Tool & Action Registry
- Security & Compliance
- Runtime Resource Management
- Monitoring & Evaluation
- Extensibility & Variation Points
- Boundary Invariants Each responsibility defines typed cross‑responsibility interactions and distinguishes condition‑dependent behaviors from replaceable implementations. To validate the architecture, we instantiate an ESG portfolio‑analysis and pre‑trade compliance agent on top of an existing service‑oriented ESG metric computation and reporting system. The case realizes seven of the ten responsibilities; different queries follow different interaction paths, controls activate only when their conditions hold, and the boundary invariants remain satisfied. We also map an independently published agentic architecture onto our reference model and find a complete correspondence without needing an extra responsibility. The results provide early evidence of the architecture’s organizational soundness and applicability, though completeness requires further study. Deployment artifacts and data are released as a replication package.
Review: The architecture offers a clear responsibility breakdown and interaction contract, giving enterprises a reusable blueprint for adopting Agentic AI, with particular strength in security and variability management.