The problem, plainly
A production agent can report success after doing the wrong thing. The response reads well. The log says done. The number in it is wrong, and a customer finds it first.
That failure is engineering-shaped. The stakes are marketing-shaped: a wrong price in a dashboard is annoying, and the same wrong price in nine thousand inboxes is a retraction.
run: SUCCESS · fact: WRONG · found by: customerWhat I'm building
The checks that sit between a demo and production. Three controls, in plain language:
Compare the actual values.
If the test already knows the correct customer ID, check it. Don't ask a second model whether the call looks right.
Stop the loop. Keep the state.
Cap the retry before it re-bills the same work. Checkpoint the run so an interruption recovers instead of starting over.
Some actions deserve a person.
Name which ones, put a real gate in front of them, and keep the trace so you can replay what happened.
None of that is impressive to watch. It's most of the code.
Proof
The first documented gate failure is in progress: a real workflow, the check that caught it, and what still breaks. It ships here, with the trace.
Campaign Loops
The newsletter: one build per issue, honestly reported, dead ends included.
Work with me
I'm open to technical roles and collaboration across AI and software engineering, mainframe and enterprise systems, platform work, and marketing engineering. The same builds documented here are the interview evidence.
tafadzwa@tafadzwam.com