Session State Continuity in Multi-Agent Voice Flows: A Systematic Review of Challenges, Patterns, and Research Gaps

Session State Continuity in Multi-Agent Voice Flows: A Systematic Review of Challenges, Patterns, and Research Gaps Multi-agent voice systems decompose telephony interactions into specialized sub-agents for authentication, billing, scheduling, and escalation. A pervasive failure in such architectures is session state loss at handoff boundaries — where caller identity, intent context, NLU slots, and operational flags silently reset when control transfers between agents. This systematic review synthesizes 36 primary sources spanning agentic AI architecture surveys, agent memory frameworks, conversational AI evaluation, voice agent deployments, and agentic AI security. Grounded in practitioner experience with Google CCAI and Dialogflow CX Next Generation Agent Studio, the paper contributes: • A four-class taxonomy of session state variables with propagation and security requirements • A classification of four failure modes (full state reset, partial propagation, stale state injection, scope isolation violation) • Platform-specific gap analysis for Google Dialogflow CX, LangGraph, OpenAI Agents SDK, and Amazon Lex • The Session State Broker (SSB) conceptual pattern for cross-agent state continuity • A six-item prioritized research agenda Keywords: multi-agent systems, conversational AI, voice agents, session state management, LLM orchestration, Dialogflow CX, CCAI, sub-agent handoff, telephony IVR, agent memory