Mmarcuscheninvoicelatency.hashnode.dev·Jun 16 · 4 min readEnd-of-speech detection is the voice-agent setting nobody tunes, and it is why yours feels slow or talks over people The lag users complain about usually is not the model thinking. It is the endpointer waiting. Two complaints kill voice agents, and they sound like opposites. "It is slow, it leaves these awkward paus00
Mmarcuscheninvoicelatency.hashnode.dev·Jun 9 · 10 min readVoice-agent observability is four jobs: 6 tools compared on tracing, eval, guardrails, and closing the loop (2026)Langfuse, Helicone, Future AGI, Arize Phoenix, Maxim AI, and Lakera across the jobs a voice agent actually needs, and why I stopped running a separate tool for each. TL;DR: "observability" for a voice00
