Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts

As AI agent systems evolve, developers are increasingly persisting procedures, constraints, and policies in natural language. While these "Agent Skills" are widely used, they have lacked a formal software engineering abstraction. The authors propose Skillware to fill this gap, treating these behavioral artifacts as first-class software objects. By defining a clear ontology, the paper provides a framework for managing these artifacts through their entire lifecycle, including discovery, activation, maintenance, and evolution.

The paper distinguishes between a Skill Artifact (the content specifying reusable behavior) and a Skillware Unit (the independently managed software identity). To qualify as Skillware, an artifact must satisfy three conditions: behavioral primacy (the natural-language source organizes the task), independent software identity (the artifact is addressable and versioned), and an execution relationship with a compatible Agent Host. The authors argue that this separation allows developers to treat natural-language instructions with the same rigor as traditional source code.

Beyond simple classification, the paper introduces an engineering lifecycle for Skillware. This includes "Identity-Preserving Evolution," a process where human contributions and agent-generated updates can be integrated into a unit without losing its software identity. By analyzing a corpus of over 138,000 Skill records and various implementation cases, the authors demonstrate that this engineering pressure is already present in the ecosystem. This framework helps move agent capabilities from ad-hoc prompts to maintainable, composable software components.

This research provides a necessary vocabulary for the emerging field of AI-native software engineering. By formalizing the boundary between the behavioral source, the agent host, and the runtime, the authors enable better governance, quality control, and interoperability for agent-based systems. It shifts the focus from merely "prompting" to building robust, versioned, and evolvable behavioral software.