Multi-agent generation framework creating extensive unbounded outputs like entire codebases or books from single input prompts. Achieves state-of-the-art generation for large codebase tasks ranking top 3 in HumanEval coding benchmark. Detects invalid code and failing unit tests with automatic error correction. Maintains complete file-store memory enabling LLM agents to read/write files. Follows exact prompt programs bypassing fixed context token limits through iterative generation. Published at ICLR 2024.