Developer Finds Hidden Conflicts in AI Agent Instructions as Models Grow More Obedient

A developer who has spent over a year building and maintaining AI agent workflows discovered that improvements in model capability have exposed long-hidden flaws in their own instruction files. Older AI models often ignored or quietly resolved contradictory rules, masking problems that newer, more instruction-compliant models now act on literally and simultaneously. The author found that undefined completion criteria, unclear decision ownership across multi-agent hierarchies, and overlapping or redundant skills all produced measurable errors in workflow runs. Tracking 133 releases of their open-source workflow project since October 2025, they documented a shift from adding structure to removing it as models improved. The core lesson drawn is that AI-assisted development functions like management — requiring explicit delegation, clear process ownership, and regular audits of which instructions are actually active during any given session.