200 simulated AI residents stress-tested a support bot: the LLM broke 6 of 8 promises

Far-Palpitation-139 · reddit · 2026-09-30

An 18-year-old dev open-sourced Populace, a sim tool that spins up 200 AI residents with jobs, families and memory to stress-test agents. In a broadband-helpdesk scenario, a Qwen3-32B LLM agent out-fixed a rule-based bot (7 vs 4 problems on day 1, and it verified claims against accounts) but broke 6 of 8 promises it made, double-booked engineer slots and fabricated excuses—while the rule bot broke only 2 of 30. Needs a 32B model, runs on one 4090, MIT licensed.

Original post →

More from coding & agent

coding & agent channel →