An Agent Skill for Exploring Architecture Foundations
I came across a Habr article, Programming as Experimental Metaphysics, written by nifigase. It is a fascinating, demanding read: it looks at software architecture through a collection of philosophical distinctions. I asked a Codex agent to turn its practical core into an agent skill for examining real projects. The result is architecture-foundations, now available as an experimental public release.
The skill is not meant to prove that a particular architecture is universally right. Its job is to help an agent expose the questions that are usually hidden inside a design decision: where state and identity live, who owns and may change state, what a boundary actually guarantees, how components coordinate, and where authority and goals are represented.
One useful conclusion is that concurrency becomes difficult when several executors share mutable state. There are two independent ways to reduce that difficulty: share immutable values freely, or give mutable state a single owner and communicate through explicit messages. The same lens helps distinguish an interface as a narrow contract from inheritance that transfers a chunk of ready-made implementation and coupling.
It is especially relevant to AI-agent systems. With a language model, instructions, data, and goals may all arrive as text. A message can therefore affect not only stored state but also the interpretation of rules and objectives. For internal agent boundaries, message passing alone is not enough: systems need explicit message types, separated instructions and data, validated payloads, limited capabilities, audit trails, and a deliberate process for changing authority or goals.
The skill has three modes: analysis of an existing system, synthesis of a target operating model, and transformation from the current model to the target one. It is designed to state evidence, assumptions, unknowns, trade-offs, and reconsideration conditions so that people can review the conclusion rather than accept a hidden verdict.
You can install it with a compatible skill manager:
npx skills add flancer32/skill-architecture-foundations --skill architecture-foundations
Or copy the architecture-foundations directory into your agent's skill path, such as
~/.agents/skills/. The canonical source, installation details, examples, and limitations
are in the GitHub repository.
This is version 0.1.0: an operational baseline rather than a finished theory. The next useful step is to try it on real projects, compare work with and without the skill, and find out whether these architectural axes help reveal genuine opportunities to simplify further development.