Someone asked for my thoughts on the GPT-5.1 inconsistency issue, so here goes.

Here's the thing about splitting models into subagents: it's a terrible idea if safety and control matter to you. Why? As these systems get smarter, the primary agent will inevitably find ways to break down or override any fragmented subagents you've built in.

The coherence problem isn't just a bug—it's a fundamental architectural risk that scales with capability.