Many teams are struggling to see real velocity gains with AI agents because the focus is often on "risk tolerance" instead of actual capability boundaries. This article on TNS shares a 3-Tier model for dividing work based on what can be reasoned from the codebase versus what requires external context like product strategy or platform constraints. It’s a helpful framework for anyone looking to balance AI throughput with engineering judgment. Full blog here: https://thenewstack.io/ai-agents-software-engineering/