Lessons From 1.9M Real Conversations Across 150 Production Sales Agents

thinkdifferent23225 · reddit · 2026-08-19

An operator running 150+ sales agents in production (WhatsApp, SMS, Instagram DMs, real customers and money) shares hard-won lessons from 1.9M conversations: every agent worked perfectly in testing — the failures only appear at real volume.

Key points:

Build terminal state gates first: an agent without a stop condition keeps selling to someone who already bought; when a "purchased/booked/not qualified" flag flips, it stops, full stop.
Agents can't reliably report their own state: instruction-based flag-setting fails silently; use a separate evaluator that reads transcripts afterward to set state. The conversational agent only controls what it says.
Follow-ups are where the revenue is: most leads don't answer the first message; one client's 1k-conversation launch included 47 conversations recovering declined payments one at a time.
Human handoff must be a designed moment, not an escape hatch: decide who gets pinged, on what channel, with what context.
The metric is booked calls, not response quality: months were spent tuning beautiful conversations that booked nothing; once scored on sales, half of "good practice" turned out to be the agent being pleasant instead of useful.
Silent failures beat loud ones: a channel showing "active" while disconnected cost a client four days of leads; alerting on zero-volume windows did more for retention than any feature.

Bottom line: they changed models twice and it moved almost nothing — the machinery around the model is the product.

Original post →

Lessons From 1.9M Real Conversations Across 150 Production Sales Agents

More from coding & agent