Reflex, Talker, Reasoner: A Tri-Tier Cognitive Architecture for Sub-200ms Full-Duplex Voice Agents with Collision Recovery

While dual-process agentic architectures such as Talker-Reasoner decouple fast conversational generation (System 1) from asynchronous multi-step planning and tool execution (System 2), they remain fundamentally bottlenecked in full-duplex spoken dialogue systems. Existing dual-tier pipelines route every acoustic pause to a generative Large Language Model (LLM) and a neural Text-to-Speech (TTS) synthesizer, incurring 350–600 ms of latency and substantial GPU compute even on trivial rejections, call terminations, or mid-sentence intent pivots. Furthermore, when aggressive End-of-Turn (EOT) silence thresholds (e.g., 200 ms) are deployed to achieve human-like responsiveness, physiological inhalation pauses frequently trigger premature agent speech. This results in mutual talk-over collisions, polite user stops, and truncated transcripts that corrupt downstream System 2 tool execution. This paper introduces the Reflex–Talker–Reasoner (RTR) architecture, a tri-tier cognitive framework that prepends a sub-cortical System 0 Reflex Layer ahead of System 1 and System 2. System 0 combines a calibrated non-generative discriminative decision gate (operating in under 85 ms) with an in-memory Pre-Baked PCM Audio Cache and a Fragile-Window Collision Recovery State Machine. On terminal rejections and conversational pivots, System 0 bypasses both the 3B-parameter generative Talker LLM and the 3B-parameter live neural TTS engine, streaming cached acoustic responses in 188 ms median latency while issuing deterministic preemption signals (CancelToken) to abort stale System 2 reasoning tasks. When a mid-breath false-EOT collision occurs within the initial 450 ms Fragile Output Window of agent playback, System 0 flushes outbound audio within 25 ms, resolves mutual-silence deadlocks via a 400 ms timer that triggers a cached yield backchannel ("Sorry, please continue"), and stitches fragmented user utterances (U1 + U2 + U3) into a single coherent turn. Evaluated on a 500-turn full-duplex tool-use and physiological breath-disfluency simulation suite, RTR reduces P50 rejection and pivot latency by 67.0%, eliminates 37.0%–38.8% of live neural TTS invocations, and improves end-to-end tool-calling accuracy under breath-disfluent speech from 33.6% to 97.4%.