weblinks

Monologue: the agent talks to itself while it listens

Waterr's Monologue runs a separate text-reasoning layer alongside a frozen real-time voice model, deliberating in the audio gaps between turns so the agent can think without adding latency. Instead of unstructured summaries it passes typed slots for verbatim facts, prior commitments and next actions, and can refresh them mid-turn while the user is still speaking. On Audio MultiChallenge this lifts a Gemini Live baseline from 38.5% to 69.8% with no change in wall-clock time, arguing that voice agents fail on what reaches the model and when, not on model capacity.