Voice bot team: GPT Live 1 and Gemini 3.8 Thinking both fail in production
Mr_Nice_ · reddit · 2026-10-23
A developer building voice bots shares production findings: GPT Live 1 delegates inconsistently — fine in simple demos, but derails in long conversations with prior delegations, and ending calls requires odd workarounds. Gemini 3.8 Extended Thinking works better but is slow and leaves dead air during long thinking pauses.
Their in-house two-agent system (a blocking tool call to a thinking endpoint, with a second agent feeding progress back as system messages) outperforms both official releases. They ask who else runs these models for complex production voice agents.
More from coding & agent
- Open-source tool turns the MacBook notch into a one-click approval panel for Claude Code — cneuralnetwork · 2026-10-23
- Taskuary: open-source bridge that routes email and chat tasks to Codex, Claude Code and Gemini CLIs — Appropriate-Path-461 · 2026-10-23
- AI makes coding 10x faster, so why isn't your team shipping 10x faster? Amdahl's law explains — arpit_bhayani · 2026-10-23
- Side project builds textual 3D worlds in the terminal with custom TUI tooling — sull · 2026-10-23
- Selfdrive AI agent turns an App Store link into a full user-acquisition plan — PrajwalTomar_ · 2026-10-23
- Thorsten Ball mocks SaaS firms building internal agent platforms nobody uses — HankYeomans · 2026-10-23