Abby Bangser argues that AI agent guardrails fail because they constrain an improvisational system instead of changing the operating environment. Writing about Kratix, she describes Promises—versioned, tested definitions that route every request, human or AI, through the same governed execution path, giving agents "bounded autonomy, not supervision." A Syntasso webinar demonstrated the approach with two agents provisioning a production PostgreSQL database. The principle: AI decides what needs doing; the Promise decides what doing it is allowed to mean.
No score is assigned. Sources and their independence are shown in the citation chain below.