Why Saving a Container Snapshot Is the Wrong Way to Preserve an AI Agent's Work

A developer building an AI agent workbench found that the intuitive solution of saving work by committing a container filesystem snapshot creates more problems than it solves. Snapshots capture unintended debris such as build caches, failed install logs, and temp files, making it impossible to distinguish useful state from noise over time. They also risk baking secrets and API tokens into reusable artifacts, turning disposable credentials into a permanent security liability. Instead of preserving the container's history, the developer argues for capturing only the declared intent: the base image, named packages, network profile, and granted directories. This minimal, human-readable recipe — called a Workshop — acts as a reproducible blueprint rather than a frozen snapshot of accidental state.