ClickHouse's Practical Approach to Agentic Coding: Beyond the Hype
Key Takeaways
- ▸Claude Sonnet 4.5 (September 2025) marked a significant quality leap, enabling more sophisticated internal tool development than earlier versions
- ▸Coding agents work best for boilerplate tasks, JavaScript applications, and internal tooling; results vary significantly by language and codebase complexity
- ▸Successful AI agent adoption requires clear use cases, proper legal/security agreements, and supporting infrastructure (observability tools, custom agents)
Summary
ClickHouse has published a candid technical case study on its real-world adoption of AI coding agents, including Claude (from Anthropic), Windsurf, and Cursor. The company began experimenting with these tools approximately one year ago, initially finding them useful for boilerplate tasks and JavaScript applications, but limited on its main C++ codebase. After signing partnerships with multiple AI vendors, ClickHouse expanded agent usage to internal tooling, performance testing dashboards, and custom agents (DWAINE, CAISER, TRAISA), coupled with AI observability platforms like Langfuse and Librechat.
A major turning point came in September 2025 with Claude Sonnet 4.5, which the company describes as a significant quality jump. This version enabled more complex internal projects, such as a Team Productivity Dashboard built in a single session with 112 prompts. However, ClickHouse's experience reveals important nuances: coding agents excel in well-defined, repetitive scenarios but remain challenged on complex, legacy C++ systems. The article emphasizes that despite the polarized discourse around AI replacing developers, the reality is more measured—agents are powerful tools for specific tasks, not universal replacements.
- Even at well-resourced tech companies, adoption remains selective: ClickHouse reports only half its team had used coding agents as of October 2025
- The practical reality contradicts both utopian and dystopian narratives—agents augment developers in specific scenarios rather than replacing engineering wholesale
Editorial Opinion
This is a refreshingly grounded analysis in an AI landscape saturated with hyperbole. ClickHouse's willingness to discuss both successes and limitations—especially the continued struggles with C++ codebases—offers more actionable insight than most vendor announcements. The article's emphasis on matching tools to specific problems rather than universal adoption reflects the pragmatism that will likely define sustainable AI integration in software development.
