Agentic Code Generation

Agentic code generation studies how language models can build software through multi-step workflows rather than producing code in a single response. Common approaches combine planning, repository or documentation retrieval, tool use, execution-based testing, and iterative debugging; multi-agent collaboration and formal verification are also used to improve complex or safety-critical work. Evaluation spans coding benchmarks such as SWE-bench and LiveCodeBench, repository-level tasks, and domain-specific tests that check whether generated systems actually run and behave as intended. A central challenge is making gains in functional correctness reliable without driving up cost or sacrificing security, maintainability, and faithful adherence to specifications.