I build software, follow questions, and write about what I find. Systems, AI, design, and the occasional loose thread.
Sameness comes from two places: training narrows what a model will produce, and a review loop narrows what survives. I've built something against both and don't yet know if either holds.
Read the essay: The Model Is a Centroid Pump. So Is the Review Loop.Most of the new advice on running coding agents validates an older idea: constraints do the work. At multi-agent scale, the constraint that matters most is who holds which decision rights.
I stopped prompting coding agents and started designing the loops that prompt them. The loops nest, tight inside medium inside long, each with its own primitives and share of human involvement.
Everyone wants to be the source of truth; almost nobody builds the conditions that let truth survive. From the Delphic Oracle to Grokipedia, the failure is structural, whatever the intent.
Every system has defaults, and AI makes them more powerful and less visible. So ask whose assumptions your system is running on.
AI can already clean-room rebuild commercial software. The useful question is how fast each kind of moat decays, and which ones regenerate faster than they erode.
Software engineering compensation was built on scarcity, not just difficulty. That scarcity is eroding. The people who survive are either unusually broad or unusually deep.
Vector search in the 90s, badges that followed you through the office, and what happens when you build an AI system without knowing what's already been built.
bristanback.com/ · Bri Stanback
