As generative AI continues to reshape the engineering landscape, the debate persists: are coding agents a boon or a bane? What once started as a controversial option for code assistance is rapidly evolving into a staple for software development teams willing to adapt and optimize their workflows. At ClickHouse, a growing consensus among engineers indicates that integrating AI agents has become essential to maintaining a competitive edge.
Distilling AI Assistance into Levels
Understanding the nuances of AI coding assistance requires a breakdown into three distinct levels of capability:
Level 1: Basic Queries. Engineers often rely on chat-based interactions with AI models to pull snippets from the internet. While this method has been employed since early days, it's largely considered antiquated when compared to more sophisticated tools.
Level 2: Integrated AI. Here, AI models are embedded within development environments, directly interacting with existing codebases. These agents can execute commands, modify code, and even commit updates autonomously, although they still require supervision for complex tasks. This is where most engineering teams are currently finding value, as it streamlines day-to-day operations.
Level 3: Autonomous Operations. This is where things get exciting but also fraught with risk. Multiple agents work collectively in a feedback loop, automating complex and isolated tasks. While this area is still in development, early adopters are starting to explore the potential for significant productivity gains.
Transformation Over Time
If you attempted to use agents a few months ago and faced failures, you've likely formed a distrust of their capabilities. That sentiment was common until significant advancements in AI tools shifted perception. A notable game-changer—Claude Opus 4.5—debuted in late 2025, which significantly improved performance, especially on complex C++ tasks. Many engineers at ClickHouse experienced newfound successes when leveraging such advanced tools, transforming initial skepticism into acceptance.
“Since Opus 4.5, agents have been usable for daily work on a large C++ codebase. 2025 was the year of the tools. 2026 should be the year of productivity gains.”
Current Applications of Coding Agents
ClickHouse's experience has demonstrated that agents find particular success in several scenarios:
Streamlining Boilerplate Work. Automating repetitive tasks—like build-system modifications or configuration changes—proves beneficial. Since agents don’t suffer from fatigue, they effectively reduce error rates in these mundane tasks.
Resolving Merge Conflicts. Agents excel at managing and resolving conflicts in code, often producing code quality superior to what human engineers might generate in high-pressure situations.
Enhanced Code Review. Implementing a proprietary bot to leverage Copilot CLI in a guided, systematic approach has shown dramatic improvements in code review efficiency. By focusing human reviewers on higher-level architectural concerns, the bot handles resource management issues, yielding cleaner final code.
Addressing Flaky Tests. ClickHouse runs a staggering amount of daily tests, often leading to a backlog of failures. AI assistance has dramatically improved this scenario, allowing developers to submit hundreds of pull requests and significantly reduce the number of ongoing issues.
Debugging Processes. While AI can identify potential bugs based on log analysis, caution is warranted. The accuracy of AI-generated hypotheses relies heavily on the engineer's own judgment, highlighting the imperative for experienced oversight.
Strategic Recommendations for AI Integration
For those looking to effectively integrate AI into their workflows, a shortlist of best practices has emerged from the past year:
- Treat AI as a tool of thought. Use these agents to enhance discussion and exploration rather than circumvent critical thinking.
- Adjust your expectations based on skill levels. Augmented capabilities will benefit strong engineers more than those lacking a foundational understanding of the codebase.
- Start with simpler tasks. Gradually build the complexity of tasks as confidence and familiarity increase.
- Prioritize validation. Invest in comprehensive testing strategies to ensure reliability, particularly when using agents for pivotal functions.
- Stay updated on AI models. Having multiple models on hand is crucial for managing downtime and ensuring consistent performance.
- Simplify guidance. Limit instruction verbosity to avoid common pitfalls associated with overly complex inputs.
- Be explicit in specifications. Clarity in task descriptions yields results more aligned with expectations.
“Treat AI as a tool of thought, not a replacement for thinking.”
Looking Ahead: What’s in Store
Although we're just scratching the surface of AI-assisted coding, upcoming developments could further revolutionize how software development functions. 2026 is expected to usher in capabilities such as autonomous bug triage and dynamic workload analysis. As these systems mature, they'll likely enhance the gap between teams that effectively leverage these tools and those that resist. For engineers willing to adapt, now is a pivotal moment to harness the fast-approaching wave of AI-driven productivity.
