Personal Agent Bench is a research website for personal AI agents: assistants designed to help with tasks such as booking, cancelling and scheduling. It combines runtime architecture teardowns with repeatable task evaluations to help readers understand how these systems work.
The architecture research examines where models run, how tools reach them, where memory lives, and what actions users approve. Readers can explore the personal agent landscape, evaluation methods, task examples, dated corrections and downloadable results.
Published completion results describe adapted runtimes under simulated exam conditions with a substituted model. They are not measurements of shipping consumer products tested end to end. Results include 95% intervals and sample sizes, and the website does not publish a combined score or ranking. This context helps readers interpret the evidence and its limits.
The website is free to access and is intended for developers, researchers and people evaluating personal AI agents. Personal Agent Bench is run by the Zentor team, which builds agent software. That creates a conflict of interest, so Zentor products stay off the scorecard.
