The industry spent the last eighteen months obsessed with “agent scaffolding”—a bloated, fragile architecture of custom-built, domain-specific wrappers that attempted to force-fit LLMs into specialized roles. It was a mistake. We were treating models like bespoke monoliths, building a new “agent” every time we needed a slightly different output.
The pivot is now clear: we are moving away from model-first intelligence toward a “code-as-interface” paradigm. The goal isn’t to build a better agent; it’s to build a better runtime.
The Death of the Bespoke Agent
The previous approach was fundamentally flawed because it ignored the difference between raw intelligence and domain expertise. You don’t want a “300 IQ mathematical genius” hallucinating tax law; you want a professional who follows a consistent, verified procedure.
Current agents are brilliant but lack institutional memory. They are stateless entities that struggle to absorb context. By shifting to “Agent Skills”—essentially organized folders of files, scripts, and markdown—we move the burden of expertise from the model’s weights to the file system.
This is the “code-as-interface” shift. Instead of baking logic into a prompt or a complex orchestration layer, we package procedural knowledge into standard, version-controlled directories. If an agent needs to style a slide deck or perform bioinformatics analysis, it doesn’t need to be “re-trained” or “re-prompted.” It simply pulls the relevant skill from the file system.
Why Files Are the Ultimate Primitive
The decision to use folders and files as the primary primitive for agent skills is a masterclass in developer experience (DX). It sidesteps the “cold start” problem inherent in traditional tool-calling.
- Self-Documentation: Scripts inside a skill folder are readable by both the model and the human. If a tool is ambiguous, the model can modify the code itself.
- Context Management: By using progressive disclosure, the model only sees the metadata (
skill.md) at runtime. It pulls the heavy logic only when necessary, keeping the context window clean and the token costs predictable. - Interoperability: Because skills are just files, they are compatible with existing workflows. You can Git-version them, zip them, or share them via standard enterprise storage.
This architecture effectively separates concerns: the Model acts as the processor, the Agent Runtime acts as the OS, and Skills act as the applications.
The Future: Continuous Learning via Procedural Memory
The most cynical take on AI agents has always been their inability to learn. Every session feels like a reset. Skills change this by providing a tangible, persistent memory layer.
When an agent writes a script to solve a recurring problem, it can save that script as a skill for its future self. This is not “learning” in the neural network sense, but it is “learning” in the software engineering sense—the accumulation of reusable, procedural assets.
We are seeing this play out in the wild. Enterprises are using skills to codify internal best practices, while developers are stitching together MCP (Model Context Protocol) servers for external data connectivity and skills for internal logic. The result is a modular, composable stack where the agent is merely the execution engine, not the repository of knowledge.
Beyond the Hype
The industry is converging on a standard architecture: an agent loop, a runtime environment, MCP for connectivity, and a library of skills for expertise.
The implications for software development are significant. We are moving toward a world where we stop “building agents” and start “curating skills.” This democratizes the process, allowing non-technical domain experts—finance, legal, HR—to contribute to the agent’s capabilities simply by structuring their knowledge into files.
If you are still building custom, monolithic agent wrappers, you are building technical debt. The future of AI implementation isn’t in the model’s reasoning capabilities; it’s in the quality of the procedural library you provide it. Stop building agents. Start building the file system they live in.