Answers are grounded in this paper and cite their sources.
Production conversational voice agents must serve many simultaneous real-time sessions without cross-session interference, resource exhaustion, or silent degradation. This paper reports concurrency and reliability lessons from a production voice platform that serves inbound and outbound business calls over SIP telephony and web channels, built on a self-hosted open-source media server with stateless agent workers, one isolated job per call, and per-agent selection of speech-to-text (STT), large language model (LLM), and text-to-speech (TTS) providers. We organize the problem around four concerns: session isolation, contention for shared and finite resource pools, admission control and overload behavior, and failure containment and cleanup. For each pattern we state whether the platform adopted it, relied on a framework default, or did not adopt it. We contribute a state-scope inventory, an inventory of the time budgets on the call path, a capacity model that allows the number of participants, streams, and requests per call to vary, and a queueing argument showing how a slow dependency converts into held capacity. We re-read ten production incidents through a concurrency lens, and we report the only concurrency evidence the deployment has: anecdotal outbound-burst failures recorded before an admission queue was introduced. No load test exists, and we make no claim about the platform's concurrency limit. We close with a load-test and fault-injection protocol, with metrics and pass criteria, as the most valuable next step. The paper is intended as a structured and candid applied systems account for engineers building or scaling similar platforms.
No takes yet. Share an insight, caveat, or question.
Altaf Rehman (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: