Agentic Coding

Quentin Chevrin, a front-end developer at Sonar, shares an advanced agentic coding workflow that directly addresses the bottleneck modern teams face: when code generation becomes cheap, verification becomes the constraint. His approach uses repository-level CLAUDE.md files to specify verification steps the agent must complete before tasks are considered done, including running tests, linters, and SonarQube analysis. This multi-layered verification system—combining deterministic SonarQube analysis, reasoning-based review through Gitar, and human review—catches different categories of issues before they reach human reviewers, allowing his squad to open four pull requests by end of day compared to the previous three-to-four days per feature. Quentin's practical PRD-driven workflow treats AI agents like junior developers: he starts with precise prompts, keeps context small by committing frequently rather than running long sessions, and calibrates scope by blast radius depending on codebase impact. He maintains control through active steering rather than parallel sessions, and intentionally writes his own Jira tickets to build the mental model needed to guide the agent effectively. His key constraints include using CLAUDE.md instructions to forbid expensive token operations (like reading full message files) and ensuring the agent never commits code directly—he does that himself to maintain visibility over changes.