AO Forge: discipline for AI delivery

AI agents write most of our production code. This is the framework that makes that a safe sentence: every feature shaped, specced, implemented and verified, with the discipline built in, not hoped for.

The idea

Agentic development fails in predictable ways: scope drift, invented requirements, silent regressions, confident claims about unverified work. AO Forge, our delivery framework, attacks each failure mode structurally, starting every feature with a requirements dialogue and ending it with a verification report that says what actually holds.

The two ideas doing the heavy lifting

Expert skills: validated navigation maps of each domain, with invariants files recording the constraints that must never break, and why. An implementing agent starts primed, not exploring. And verification as a first-class phase: spec-derived tests gate the work, and reviewer-facing walkthroughs tour every diff with the spec's intent in hand.

The proof

AOne, the CMS publishing this site, was built through this framework: security hardening, three-environment promotion, publish channels, a full site-building UI. When reality disagreed with the plan, the record says so. Honest deviations are part of the method.