User-written permission policies: human approval leads in preventing AI agent overreach - a 113-participant study

Do reusable allow/ask/never rules - permission policies written in advance by the user - protect better against agent overreach than runtime review? That is the central question of a study by Ting Yan, published as an arXiv preprint on August 27, 2026, with no institutional affiliation listed on the arXiv record.

In a controlled study with 113 participants, simulated overreach events by agents running through an MCP proxy were tested under three conditions: user-authored policies, human approval for each action, and automated review.

The user-authored policy condition prevented 20.1 percentage points less simulated overreach than per-action human approval, and 14.5 points less than automated review. The study's advice is that approval design must be consequence-aware.

The practical recommendation: persist workflow state, idempotency keys, tool inputs, policy decisions, approvals, outputs and compensating actions - recovery and audit depend on it.

Limitations: a simulated single-day setting, a single participant pool and a single policy design; the author notes it is not a universal comparison across all permission systems.