The author builds a code generation factory using multiple AI agents (claude code, o3, sonnet 4, sonnet 3.7) each in its own git worktree to plan, execute, verify, and improve code automatically.
The guiding principle is to fix inputs—plans, prompts, and agent configurations—rather than hand-patch outputs to ensure reliability by construction.
Daily workflow involves two steps: planning with o3 to generate detailed plans and execution with sonnet agents that write commits, verify code against plans and original requests, and feed issues back into the plan templates.
The system runs parallel agent attempts, iteratively improving inputs so the factory becomes self-improving and reduces manual intervention.
Scaling is achieved by creating specialized agents for style enforcement and library integration, and by composing small agents to automate complex workflows like API integration, tests, and documentation.
Future enhancements include better coordination of agents, aligning business documentation with agent inputs, building more complex workflows, and optimizing token usage across AI providers.
Get notified when new stories are published for "General AI News"