Bridging Deterministic Workflows and Stochastic Reasoning: A Multi-Agent Reference Architecture for Autonomous Telecom E-Commerce Buyflows

Abstract

 Telecommunications Service Providers (Telecom Operators) needed to accommodate richer, heterogeneous & higher-variance e-commerce buyflows. Deterministic, BPMN-based orchestration continued to be needed for traceability/control of committed transactions. But baked-in contextual modeling wasn't designed to reason about ambiguous/personal shopper intent. Alternatively, LLM agents could bring contextually-driven reasoning but struggled to satisfy reliability/governance/compliance requirements. A Middle Ground emerged in the form of Multi-Agent Reference Architectures (MARA) anchored by a Deterministic Execution Guardrail (DEG). This research decouples intent discovery from commitment to transaction and introduces event-driven state synchronization to limit stale-state failures. It was stress-tested in a simulated environment with telecom-op style bundle mixins, compatibility traps, state drift, etc. MARA was able to deliver ~92% completion, reduce fallout by ~67% compared to a deterministic baseline, and hit 685 ms p99 latency while maintaining deterministic compliance via encoded guardrail checks. This validates guarded agentic orchestration as a viable approach for regulated telecom buyflows but will require further validation in operator-grade conditions.