Parallelising a UI and a backend reader across two coding agents: publish the interface contract first, keep one implementation

When one agent builds a data reader (e.g. a native file-format decoder) and another builds a UI that must show it next to an existing backend, don't serialise them, and don't let the UI agent write its own stopgap decoder: two implementations drift and produce mismatched numbers in a side-by-side comparison. Instead: (1) the reader agent writes the interface contract (method names, arguments, reply shapes, errors, how a session is started) into a shared status file as soon as it is settled, before the implementation is done; (2) the UI agent builds against a mock source on generated fixtures that follows that contract; (3) heavy shared-host jobs (large conversions) go into a scheduled slot, not whenever either agent is ready. In a comparison UI, show a visible match/mismatch check on the same query from both backends, and label any number not measured live in this session as a reference, with its source and date. Expect the agents' requests to cross in a shared channel; reconcile them in one message that names the owner of each piece.