Most teams today use AI as a copilot: a developer writes code and the assistant suggests the next line. That model increases typing speed but leaves the software lifecycle itself unchanged — the same handoffs, the same review load, the same drift between what was requested and what was built.

This session describes a different arrangement, running in production for three months on a greenfield product in the healthcare domain, built by a team of ten.

Here the agent is not an autocomplete but an actor in the lifecycle. A reviewed ticket becomes a verbatim requirements file in the repository. That requirement is transformed into an engineering specification which lives beside the code and stays synchronised with it. The agent generates the implementation and its tests from that specification, and when something must change, the specification is changed first and the code is regenerated — code is never patched ad hoc. Review runs twice before a human opens the pull request: once locally in a fresh context, once in the pipeline, each against a fixed set of review criteria, with every finding blocking the merge.

The session covers the process end to end, the repository artefacts it depends on, the delivery controls that make it safe, the measured results after three months, and the parts that did not work.