AI Will Solve the Problem. It Will Not Protect Your Architecture.

I was setting up a website fully inside Cloudflare. Strict constraints. No third party tools. No external services. Clean stack. Minimal surface area. Reliability first.

I had everything structured clearly. Skills defined. Phased build plan. Tight instructions.

What I noticed was interesting.

Even with explicit constraints, Claude kept suggesting external services as “easy” or “free” solutions. Things like adding a mail service or hooking in another tool to solve a small gap.

Technically correct. But strategically wrong for this build.

I had to shut it down multiple times and restate the boundary:

We are not adding third party services.

Everything stays inside Cloudflare.

No new accounts. No new APIs. No new failure points.

The drift wasn’t malicious. It was trying to be helpful. It saw a solution and offered it. But that solution multiplied complexity and risk.

That reinforced something for me:

AI will optimize for solving the problem.

It will not automatically optimize for architectural restraint.

If you are not explicit and persistent, the stack will sprawl.

This is where human-in-the-loop really matters. Especially when you care about:

• Risk mitigation

• Maintenance overhead

• Control of dependencies

• Long term reliability

The model can solve the issue. But it won’t necessarily solve it in the way you would architect it.

So my takeaway:

When building with strict constraints, watch for drift. Even with strong prompts and structured phases, it will try to “improve” the system unless you enforce the boundary.

Progress overall is solid. Just an interesting observation about agentic behavior in real-world builds.

8 comments