Read attributed personal agent field reports. Each report covers a particular environment and outcome, not a general success rate.
One information-organizing run on five previously researched sources produced 20 initial claims and 25 revised claims. AI review corrected wording, disclosures, a missing log event and the attribution of human input. The user confirmed the content; the source tasks were not reproduced.
A runnable Node example fetches feeds, removes duplicate URLs and remembers seen items. Live inputs produced 35 unique entries; an unchanged replay produced zero new entries. Agent scheduling remains untested.
Saurabh Das describes a Raspberry Pi agent connected to Telegram and cloud models for market, weather, astronomy and chess alerts. He reports easier maintenance and lower token overhead after moving from OpenClaw, without reproducible comparative measurements. Treat this as personal experience.