The landscape of software development is shifting beneath our feet, and Emanuele Pugliese, founder of PuglieseWeb, has a provocative thesis: AI has effectively promoted every developer on your team, but most haven’t updated their job descriptions. In a recent blog post, Pugliese argues that while AI agents have taken over the mechanical aspects of production—writing scenarios, drafting documentation, and generating code—the human role has migrated to high-level architectural judgment. This shift necessitates a new practice he calls Architecture as Truth (AAT), where versioned documents become the single source of authority for both human teams and AI agents.

The End of Silent Production

Pugliese illustrates this shift with the case of "Marta," an analyst whose role in writing acceptance scenarios was reduced from two days to minutes by AI. However, the AI produced three scenarios that were technically correct but contextually wrong for the company’s specific financial rules regarding credit notes versus refunds. This highlights the core problem: AI excels at general production but fails at specific, historical, and contractual knowledge that lives only in the heads of senior staff. The new "entry requirement" for developers is no longer just coding skill, but a deep, end-to-end understanding of the product’s promises and constraints.

Defining Architecture as Truth

AAT is defined as a methodology where architecture documents—decisions, contracts, and behaviors—are the definitive source of truth. People approve changes to these documents, not just code diffs, and AI agents build strictly from this approved truth. The practice rests on five principles, including "Write the truth down" and "Hold the system to the truth." Unlike standard spec-driven development, where specs are often retired after a feature ships, AAT maintains standing documents that are continuously reconciled with the codebase. This ensures that the system never drifts quietly from its intended design.

SystemDox and the Agent Contract

To implement this at scale, PuglieseWeb is developing SystemDox, a tool currently in private preview that serves these architecture documents to AI agents via the Model Context Protocol (MCP). The tool enforces a strict contract for agents: read before writing, build only what was approved, stop and ask when the truth is silent, and leave the truth updated. In one real-world example, an agent working on a reminder feature was able to identify a conflict with an existing decision record (ADR-214) and propose an amendment, preventing the accidental creation of a duplicate reminder system. This demonstrates how structured context can prevent AI hallucinations in complex codebases.

Key Takeaways

  • AI has commoditized code production, shifting the developer's value to architectural judgment and product understanding.
  • Architecture as Truth (AAT) positions versioned documents as the single source of truth for both humans and AI agents.
  • SystemDox, a private preview tool, uses MCP to serve architectural context to coding agents, enforcing strict read-before-write rules.
  • The methodology requires that agents stop and request human approval when architectural truth is ambiguous or missing.

The Bottom Line

This isn't just about better prompts; it's about treating architecture as a living, versioned database that AI must query before it acts. If your team isn't writing down the "why" and "what" in a format agents can read, you're just guessing at scale.