ClickHouse's CTO shares a detailed, experience-driven account of how the company uses AI coding agents in production. The post covers the evolution from skepticism to full adoption, specific use cases including fixing flaky tests (700 PRs in two months), investigating complex bugs, code reviews, porting code between codebases, and vibe-coding internal tools. It introduces a three-level framework for AI-assisted coding, compares CLI agents (Claude Code, Codex) vs IDE integrations, and provides practical recommendations like maintaining CLAUDE.md files, running parallel sessions, and always validating agent output. The post also addresses common fears around AI adoption — job loss, skill atrophy, quality degradation — with candid, grounded responses. A key result: flaky test failures dropped from ~200/day to 3–5/day per 10M tests using agent-assisted fixes.
Table of contents
Safe assumptions #Why now? #It's easy to be skeptical #Levels of AI coding #Available tools #Usage scenarios at ClickHouse #Usage recommendations #AI in open-source #AI FUD checklist #We should use it more #Sort: