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

A new C program, approximately 1000 lines long, has been released that allows an LLM (large language model) to rewire its directed acyclic graphs (DAGs) in real-time during execution. This innovative runtime environment enables an LLM to adapt the workflow based on error feedback; when a task fails, the LLM receives the error output and can choose to RETRY, PATCH, INSERT_BEFORE, or ABORT the task. The program maintains a detailed log of changes, allowing users to track precisely how the LLM altered the execution graph and why certain decisions were made. This development is significant for the AI/ML community as it exemplifies a novel approach to interactive machine learning workflows, promoting enhanced flexibility and resilience in automated processes. The system supports YAML-defined tasks and runs on various operating systems, showcasing its accessibility. Importantly, the LLM operates as a peer to the scheduler rather than an autonomous agent, and it is designed to make limited modifications to the graph based on task performance, ensuring sophisticated error handling while preventing disruptive changes. By emphasizing a structured logging system and constrained decision-making, this program sets a new standard for integrating LLMs with traditional programming paradigms.