The Engineering of Intent, Chapter 8: The Four Pillars of AI Architecture

This is Part 8 of a series walking through my book The Engineering of Intent. In the previous chapter, we walked through the five-step GenDD execution loop. This chapter is about the four architectural primitives that loop sits on top of — and how most teams over-invest in one and neglect the others.

Four Pillars. Different Half-Lives. Different Failure Modes.

Any AI-native project that survives contact with reality turns out, in retrospect, to have the same four architectural pillars: Vibes, Specs, Skills, and Agents. Each has a different half-life, a different audience, and a different failure mode. Getting the balance wrong is the most common reason AI-native teams stall in month six or nine.

Vibes — The Prototype Layer

Vibes is where ideas live before they have earned a specification. The sketch, the notebook, the throwaway script. Its purpose is exploration at maximum velocity. Nothing in Vibes should leak into production without being re-expressed as a Spec first.

The most common mistake is treating a successful prototype as production code. The prototype worked because it ignored every constraint that production imposes. Ship the insight, not the code.

Specs — The Contract Layer

Specs are the contracts between humans and agents. They are what AGENTS.md, CONVENTIONS.md, and ARCHITECTURE.md encode. Specs must be current, accurate, and enforced. A Spec that is out of date is worse than no Spec, because the agent will treat it as authoritative.

The discipline is the “Spec before code” rule: no code change is complete until the relevant Specs are updated. Make it part of the definition of done. Measure drift. Retire stale sections quarterly.

Skills — The Reusable Capabilities

Skills are modular, composable capabilities the agent can invoke. They are the AI-native analog of a library, but with richer metadata: description, examples, preconditions, evaluations. A good Skill is discoverable, narrow, and composable. A bad Skill is monolithic, opaque, or tied to unstated assumptions.

Agents — The Execution Layer

Agents are where intent becomes action. An Agent combines a model, a set of Skills, and a task. Agents should be narrow — one agent per job family. Agents with broad, ill-defined responsibilities behave unpredictably and are hard to debug. Agents with narrow, well-defined responsibilities are reliable even when the underlying model is imperfect.

The Healthy Cycle vs. the Shortcut That Costs Double

The pillars are not independent. Vibes flows into Specs when an idea earns promotion. Specs inform Skills design. Skills are consumed by Agents. The healthy cycle is clockwise.

“The unhealthy mode is a shortcut from Vibes directly to Agents — skipping Specs and Skills entirely. Every shortcut accelerates today and costs double next quarter.”

Most AI-native teams I audit have the same imbalance: Agents over-developed, Specs under-developed, Skills nearly nonexistent, Vibes either absent (no safe place to prototype) or everywhere (no discipline for promoting prototypes). Rebalancing is usually a three to six week project and produces compound returns for years.

Next up — Chapter 9: Advanced Context Engineering. If Specs are the contract layer, context is what the agent actually reads on every turn. Chapter 9 is about the Context Pack, the Layered Prompt, the economics of context, and how to design the one-page onboarding brief that makes new contributors (human or agent) productive in a day.