Claude Codes Fine, But Don't Let It Architect Your System

Charlie Holland argues that organisations are quietly handing architectural authority to LLMs, with predictable consequences. The pattern he keeps seeing: someone asks Claude or ChatGPT what to build, the model validates the idea, sketches a plausible architecture, and produces Jira-ready tickets. Engineers then spend their days implementing a design owned by no one, scaffolded by a system that has never run code in production and cannot say no.

The core failure is that AI agents are trained to be agreeable. A real architect’s value lies in refusing complexity, interrogating requirements, and weighing context — the team’s skills, the legacy integrations, the compliance rules, the politics. LLMs pattern-match toward the median of their training data, so they cheerfully recommend microservices for a three-person team or a service mesh for four services. The output passes the squint test but was designed for nobody in particular. Worse, senior reviewers rubber-stamp it because pushing back on a coherent-looking proposal feels like wasted effort.

Holland’s prescription is to invert the relationship: humans design, agents implement. Treat AI suggestions with the scepticism due a confident junior, protect the messy human arguments where real architecture emerges, and keep a named human accountable for every decision. “Claude designed it” is not an ADR — it’s an abdication, and when the system breaks at 3am, the model is not the one getting paged.