Lessons from running AI agents on 60k real estate leads: smaller prompts, more code
Warm-Reaction-456 · reddit · 2026-10-04
A team shares what changed after months of running text and voice agents on a 60k-lead Canadian real estate database:
- Shorter prompt, more code: v1 put opt-out and wrong-number rules in the prompt and they kept breaking. Now anything with a right answer lives in plain code; the model only reads replies and writes the next message.
- Intent labels collapsed to four, one per action the realtor actually takes (call now, follow up by date, keep warm, drop) — labels that don't map to a human action are trivia.
- Weird replies dominate: late replies against old threads, "who is this?", family members answering for people who moved or died — scripts decided in advance.
- Voice lessons: one sentence per turn, stop when the human speaks; voicemail copy matters more than expected; text first, voice only for non-repliers since minutes are expensive.
- A human reads random conversations weekly — dashboards can't catch the agent getting pushy or misreading a reply type.
- Every weird conversation goes into a replayable test set, and model versions are pinned.
Summary: production made the agent smaller and dumber and everything around it bigger — and it works far better.
Related event: Real-World Lessons From Running a Real Estate Lead Agent for Months(2 posts)→
More from coding & agent
- Local LLM benchmarks are mostly noise: c=1 tokens/s hides real concurrency performance — TheZachMueller · 2026-10-04
- Even Opus 5.5 ships vulnerabilities when your vibe coding requirements are vague — gefei55 · 2026-10-04
- Too many agents to track: David Khourshid loses track of what his own bots do — DavidKPiano · 2026-10-04
- Harness engineering reduced to prompts and tools, but it's what makes agents reliable — techNmak · 2026-10-04
- A model alone isn't an agent: a first-principles handbook on agent harness engineering — techNmak · 2026-10-04
- Building a Miro clone 5x on 3 local rigs: tokens/sec is useless, thinking variance hits 5x — julianharris · 2026-10-04