The team says it built a standalone Facebook Groups app with a small team, and measured success by verified, production-ready changes rather than code volume. Agentic delivery depends less on generating code than on making completion observable and reviewable.
The mechanism is: A closed plan-build-test-correct cycle uses automated checks and recorded evidence before human review, with engineers retaining ownership of the goal and approval. Meta reports integrating build status, unit tests, and UI-spec matching, then attaching test plans and recorded verification sessions to pull requests. It also equipped navigation agents to exercise the live app and inspect screen state rather than guess at coordinates. Meta says engineers still approve shipped outcomes, and some stages still keep an engineer in each iteration. This is a company-authored account of one product effort, not an independent measurement or proof that the workflow generalizes.
The report makes clear that moving engineers out of every iteration shifts the constraint to trustworthy verification, not simply faster code production. Teams adopting coding agents need executable checks and evidence attached to changes to avoid increasing review burden. Whether more stages can run autonomously while engineers retain sign-off based on reliable build, test, and UI evidence. Meta frames human-on-the-loop as a target, not a completed state.