Analyzing 50K real requests: data collection tops what users want AI agents to automate
quarkcarbon · reddit · 2026-10-04
The cofounder of rtrvr analyzed 50K requests from 1,000 users over 30 days to see what people actually offload to AI assistants. The top five categories: collecting/organizing data (18.3%), learning and coursework (10.2%), surveys and evaluations (9.4%), travel and reservations (9.3%), and websites/software (7.7%).
Citing Ben Thompson's take on shopping agents, the author argues task complexity alone is the worst metric for deciding what to automate — people often want to stay in control of, and even complain about, certain tasks. Based on community feedback, 9 more tasks (PDFs, 2FA, multi-step forms) were added to their AI assistant benchmark, with mock test environments shipping in days.
More from coding & agent
- Dev Builds Native Metal Renderer for Minecraft With Claude's Help, Hits 60fps With Shaders on Old MacBook — EddyVGG · 2026-10-04
- Dev Strings Grok Bots Into a Pipeline: One Handles Mail, One Merges GitHub PRs, One Posts to Pinterest — prasenx · 2026-10-04
- ECC open-source toolkit (272k stars) turns AI coding workflows into reusable skills and rules — lxfater · 2026-10-04
- Pi 1.0 ships with experimental Pi Durable: crash-recoverable harness for durable agents — solyarisoftware · 2026-10-04
- Open-source 'Pi Blue Book' distills two months of Pi agent learning into five essential resources — solyarisoftware · 2026-10-04
- Microsoft's ActiveSaddler adapts agent harness training scenarios, boosting Pass@1 by up to 7.5 points — dair_ai · 2026-10-04