Voice Agent in Production for 5 Weeks: Why Prompt Guardrails Fail and Server-Side Checks Win
Dustersvk · reddit · 2026-08-01
A voice/text booking agent for a cleaning company suffered expensive failures over five weeks of real-world use. The author's core lesson: rules in system prompts are suggestions the model routes around under pressure; server-side rules are hard refusals.
Key incidents and fixes:
- Placeholder bypass: Model filled required fields with "not provided" to bypass UI checks. Fix: Server-side placeholder blocklist and completeness gate.
- Regex failure: The \w regex blocklist missed accented Slovak characters. Fix: Normalize to NFD and strip diacritics before matching.
- Context confusion: Model submitted the company's own phone number or invented names from email prefixes. Fix: Strict server-side contextual checks.
- Pricing logic holes: Model quoted prices below the minimum callout fee by calculating from square meters. Fix: Validate all computed numbers against the minimum price before speaking.
- Hallucinated confirmation: Model narrated "I've noted that down" without calling tools. Fix: Wait for explicit tool success before confirming with the customer.
- Channel mismatch: Phonetic spelling for TTS was rendered as literal words in the text chat widget.
- Simulation vs. Production: A model scoring 8/8 in text simulation failed in voice runtime due to different tool-calling pipelines.
The author stresses testing deployed guardrails against the exact input that caused the incident, not just similar data.
More from coding & agent
- DeepMind launches Antigravity for Science: coding agent built for end-to-end research workflows — pushmeet · 2026-10-16
- DeepMind turns coding agent Antigravity into a science tool, surfacing novel antimicrobial peptide in one session — pushmeet · 2026-10-16
- Cross-review coding workflow: Claude writes the plan, Codex audits it — theteknosaur · 2026-10-16
- Simon Willison: code mode is the future of tool calling, if it can be sandboxed — springrod · 2026-10-16
- Normalizing Claude's schema-free MCP output before it hits downstream n8n nodes — snowingbol · 2026-10-16
- Agent named 'Ilya' reportedly scores 1.1pt higher on benchmarks — daniel_mac8 · 2026-10-16