Building voice agents that talk while they work: JOSIE-Voice's tool-calling playbook
solyarisoftware · x · 2026-10-19
JOSIE-Voice shares engineering patterns for tool calling in voice calls, where a spinner is silence and silence feels broken:
- Speak before the tool, not after: utterances like "Let me check your calendar" are spoken the moment the tool call starts, not glued to the result.
- Short, bounded updates: a filler line before the first step, a progress line every 12 s, max 3 asides per turn.
- Parallel tools with deadlines: up to 5 tool rounds per turn, parallel calls, 15 s deadline per call (longer for Shortcuts, Music, MCP); interrupting cancels running tools.
- Conversation doesn't wait: sub-agents keep running in the background.
More from coding & agent
- DeepSeek opens Search API: tester spends $0.08 on 110 calls, finds authority scores tied to hosts — gaganghotra_ · 2026-10-19
- DeepSeek-V4.1-Flash-1017 shown building full projects from scratch — FinanceYF5 · 2026-10-19
- DHH says he routes all Claude Code traffic over Tailscale while in China — op7418 · 2026-10-19
- Why JavaScript is the best investment for AI/ML engineers building agents — kmeanskaran · 2026-10-19
- Ramp's Glass AI workspace bets on team-wide sharing of reusable agent skills — reitiux · 2026-10-19
- LiveMACEBench: process-aware eval of LLM agents in live financial markets — Fudan-University · 2026-10-19