Prasenjit Sarkar is a solutions marketing manager at Sonar focused on enterprise-quality AI-generated code: whether software produced by coding agents remains secure, maintainable and affordable after it passes its tests. His background spans more than two decades in technology and product work, including cloud-native networking, API management, language models, retrieval-augmented generation and knowledge graphs.
In 2025, Sarkar evaluated four GPT-5 reasoning settings across more than 4,400 Java programming assignments. Greater reasoning intensity improved functional performance and reduced some obvious defects, but also produced more code and shifted failures toward subtler issues, including concurrency errors and incomplete certificate validation. The medium setting slightly surpassed the high setting on functional success while costing substantially less.
He subsequently coauthored research comparing six matched pairs of cleaner and messier repositories with Olivier Schmitt and Priyansh Trivedi. Agents working in cleaner codebases used approximately 7% fewer input tokens and 8% fewer output tokens, and revisited edited files less frequently—evidence that software maintainability affects operating costs as well as developer comprehension.
- Passing tests does not establish production readiness. At AI Engineer Europe 2026, Sarkar described evaluating 4,444 Java assignments for security findings, bugs, code volume, cyclomatic complexity and cognitive complexity. Models solving identical problems generated markedly different amounts of code: Claude Sonnet 4.6 produced approximately 627,000 lines, while GPT-5.4 variants produced approximately 1.2 million.
- Clean code changes agent economics. Repository structure, naming and manageable complexity influence how much context coding agents consume and how often they retrace their work, making technical debt a recurring computational expense.
- Deterministic verification makes autonomy accountable. Sarkar advocates reproducible verification gates grounded in compilation, static analysis and security rules. Probabilistic model-based review can assess intent, but enterprise workflows also need objective checks capable of rejecting unsafe or incomplete changes.
- Guide, verify and solve form one development cycle. His Agent-Centric Development Cycle pairs architectural context with in-loop code analysis and verified remediation while retaining developer oversight. He also coauthored the introduction of Sonar Vortex and the SonarQube Remediation Agent, extending those principles across code generation, quality gates and automated fixes.
