A cascading voice agent chains speech-to-text, an LLM, and text-to-speech — and every layer downstream can only react to the transcript it's handed. So your real accuracy ceiling is set at the STT layer, not the LLM. This piece maps the four places context lives in the stack, shows how to feed the agent's own turns back into Universal-3.5 Pro Realtime with agent_context on LiveKit and Pipecat, and makes the case that most teams pick their STT last when they should pick it first.