AI Wars Ep. 3: The Return of the Developer

A developer at Marmelab used Claude Code to build a COBOL admin interface generator from OpenAPI specs, despite having no prior COBOL experience. The experiment revealed key limitations of AI coding agents: they generate insecure code by default (path traversal, XSS, command injection via CALL SYSTEM), fail to identify root causes when fixing bugs, make poor architectural trade-offs without weighing risks, and produce bland UIs. The author argues that developers remain essential for architectural decisions, detecting when AI goes off-rails, enforcing good practices, and taking responsibility for the output. Automated tests proved critical for enabling the AI to iterate independently.