What they did
Rangel directs and verifies but does not write the code. For a change to a KPI and payroll platform that swapped native browser tooltips for the app's own component, he set up Claude Code with file-defined agents, one per role, each with restricted tools. A test agent writes one failing test per behaviour, with the expected value worked out by hand from the work card rather than copied from program output. A builder agent writes the least code that makes that test pass and may not edit tests. Rangel runs every test himself and removes the fix to confirm the test goes red again. The loop ran eight times. Last, a read-only reviewer agent in a clean session got only the original card and the final diff, and checked it against seven fixed questions: scope, off-limits areas, whether the test would fail without the fix, invented business rules, hidden failures, UI standards and screen states.
How it went
Native tooltips on the running platform went from 186 to 0, with 46 screenshots in light and dark. Counting by hand showed the card's figure of 15 tooltips was really 17. On the final step nine old tests broke, and the builder stopped and reported them instead of patching them. The reviewer made eight findings, two of them blockers: a folder of browser tests nobody had opened still used the removed attribute, and the plan contradicted the code on keyboard focus. It also caught a missing performance check; opening a week now takes 265 ms instead of 152 ms.
Worth knowing
The reviewer is useful because it knows nothing about how the change was built. Rangel had verified every step himself and still missed both blockers, since he only looked where his plan told him to. The roles are enforced by tool limits in the agent files, not only by instructions.