What ClickHouse learned from a year of coding with AI agents

ClickHouse CTO Alexey Milovidov shares lessons from a year of using AI coding agents on a massive C++ codebase. He outlines three levels of AI-assisted coding and explains how Claude Opus 4.5 was the turning point that made agents genuinely useful for daily C++ work. Key wins include resolving merge conflicts, automated code review, fixing flaky tests at scale (reducing CI failures from ~200/day to 3-5 per 10M runs), and debugging hard concurrency bugs. Seven practical recommendations are offered: treat AI as a thought tool not a replacement, recognize it as a multiplier of existing skill, start small, always validate with tests, use latest models from multiple providers, keep guidance files short, and be specific in prompts. Autonomous Level 3 agents are identified as the next frontier.