Skip to content

Will AI bypass the boundary to get it working?

Elevator answer

Yes—an assistant can make the test green by injecting an HTTP client straight into a business step. Typed connector rules and review must make that shortcut loud.

3 coffee-machine misconceptions

AI Ada

“It compiled, therefore the boundary is valid.”

Framework Fred

“A direct client call is only temporary.”

Enterprise Edna

“Prompt the model not to leak data.”

The real explanation

Ask an assistant to “check the customer's risk score” and the shortest solution may be riskClient.get() inside ApproveOrderStep. The feature works locally. Retry identity, credentials, capture, telemetry, and ownership have quietly vanished into a method call. AI did not invent that shortcut; it can merely type it before the architect finishes inhaling.

If the assistant places stripeClient.charge() inside CalculateTotal, the declared flow has no matching Command or connector operation. That gives the compiler, review tooling, and a human something objective to reject. The problem is no longer “I dislike this style”; it is “this effect has no identity, recovery rule, or owner.”

The framework cannot prevent bad changes by itself. Teams need guardrails, reviews, tests, and a culture that treats generated suggestions as proposals. AI is most useful when it helps implement a clear contract, not when it invents a new integration path because that was faster.

Trade-offs

TPF gains visible safety rails. It gives up some shortcut convenience.

When TPF is not a good fit

If teams routinely accept unreviewed generated changes in consequential flows, first solve that delivery practice.