Architects Who Cannot Do Architecture - How AI Agents Expose the Gap
Architects Who Cannot Do Architecture - How AI Agents Expose the Gap
There is a common pattern in tech companies: someone draws a diagram with boxes and arrows, labels it "architecture," and calls the job done. The boxes say things like "API Gateway" and "Event Bus" and "ML Pipeline." The arrows connect them in satisfying ways. But nobody has verified that this design actually works.
AI agents are exposing this gap in a way that is hard to ignore.
The Diagram-to-Reality Gap
When you ask an AI agent to implement an architecture diagram, it takes the specification literally. It tries to build exactly what the diagram describes. And that is when the problems surface:
- The "Event Bus" box has no defined message format
- The arrows between services have no error handling specification
- The "ML Pipeline" assumes data arrives clean and pre-formatted
- The authentication flow has three different approaches drawn on the same diagram
- Scaling is described as "auto-scale" with no actual mechanism
The agent cannot fill in these gaps with tribal knowledge or hallway conversations. It builds what was specified. And what was specified is incomplete.
Real Architecture Is Executable
Good architecture survives contact with implementation. The test is simple: can someone - human or agent - take your architecture document and build a working system from it? If the answer requires "well, you also need to know that..." then the architecture is incomplete.
AI agents are ruthless about this. They do not make assumptions. They do not ask clarifying questions in most cases. They build what the document says and the result reveals every gap.
Using Agents as Architecture Validators
This is actually a useful pattern. Before committing to an architecture, have an AI agent attempt a prototype implementation from the design document alone. The places where the agent gets stuck or makes wrong assumptions are exactly the places where your architecture is underspecified.
It is cheaper to find these gaps with a prototype agent than to discover them six months into development with a full team.
Fazm is an open source macOS AI agent. Open source on GitHub.