Multi-agent handoffs during live voice calls lose conversation state because in-memory and database-backed frameworks are architecturally incompatible

Production voice applications increasingly need mid-call agent handoffs -- routing from a triage agent to a specialist, escalating from bot to human, or switching personas without dropping the call. Pipecat's in-memory session model is optimized for low-latency streaming and is fundamentally incompatible with agentic frameworks like CrewAI and LangChain that rely on database-backed session persistence. When a handoff occurs, the receiving agent starts cold with no conversation state, no tool call history, and no awareness of what the caller already said. The Pipecat GitHub issue #2763 (October 2025) documents this gap: developers cannot do seamless multi-agent handoffs in an active call without rebuilding session state from scratch. The HN thread on the OpenAI gpt-realtime launch for SIP telephony shows further evidence: the new /accept/ endpoint has zero documented parameters, handoffs across SIP legs are undocumented, and developers who need specialist routing during a call have no reference implementation. This is distinct from the MCP server multi-agent coordination (artifact versioning / write locks) already in the DB -- that signal is about coding agents sharing files. This signal is the real-time voice-specific handoff problem where latency budget (sub-500ms) and in-memory architecture preclude normal distributed state approaches.