The software factory is structured as two nested loops: an inner loop that executes tasks (e.g., a triage agent classifying issues using skills) and an outer loop that improves the system itself by reviewing feedback (e.g., comments) and rewriting the agent skills. The unit of agent capability is a 'Skill MD file' that the outer loop mutates.
Why it matters
Places the self-improvement mechanism at an explicit architectural layer, separate from task execution. It implies agent skills are first-class, versioned artifacts produced by an automated process rather than hand-authored prompts, and that a harness must expose feedback channels (comments, review outcomes) back into skill generation.
Generalization
Any long-running agent system can be split into an execution plane and a separate improvement plane that consumes execution telemetry to rewrite its own capability definitions—making 'self-improving' an addressable component rather than a property.
Inner loop handles task execution (e.g., triage agent classifying issues using skills), while outer loop handles system improvement (e.g., skill updater reviewing comments and improving agent skills).
An effective factory requires automation, context and skills, human-in-the-loop triggers, and self-improvement loops.