We use ClickHouse for a trend detection pipeline and the agentic coding part resonates. Half my codebase was written with Claude Code at this point. The tricky bit is when the agent makes changes that look right but break some subtle assumption three files away. Curious how they handle that at scale with a codebase the size of ClickHouse.