Show HN: A tiny C program where an LLM rewires its DAG while running (github.com)

A newly developed C program, approximately 1000 lines long, introduces a notable innovation in running YAML-defined Directed Acyclic Graphs (DAGs) by allowing a language model (LLM) to modify the graph while it is executing. When a task fails, the system invokes a planning LLM that analyzes the error output and suggests one of four actions: RETRY, PATCH, INSERT_BEFORE, or ABORT. These modifications are then applied dynamically, enabling the program to adapt in real time. An append-only event log records all changes, making it possible to replay any execution and trace the LLM’s decision-making process. This development is significant for the AI/ML community as it highlights a novel approach to integrating LLMs within operational workflows, positioning the LLM as a collaborative tool alongside the task scheduler rather than a standalone agent. The implementation requires no third-party libraries and is compatible with Linux, macOS, and Windows, utilizing a simple CLI to execute tasks and capture their states. With features like task-specific mutation budgets and clear delineation of task origins, the program not only enhances fault tolerance but also improves the clarity of task management within workflows, potentially paving the way for more intelligent and adaptive software systems.