Engineering in
the agentic era.

The automated, per-PR process of checking whether a code change conforms to the approved architectural intent of the system — run at CI/CD speed, every time a PR is opened.

The property of a software system whose architectural intent is known, whose changes are continuously measured against that intent, and whose audit is automated at CI/CD speed.

Every CI/CD pipeline has code quality gates. None have design intent gates. The evolved pipeline adds the one gate that governs AI-generated code at scale.

The only way to implement SecDevOps when AI agents are generating production code is through enforced design intent at every PR — automated, continuous, and not dependent on an architect being present.

The silent, incremental divergence between the architectural intent a team agreed to and the implementation that actually ships.

Two distinct governance mechanisms that answer different questions: code review asks 'does this code pass?' — design review asks 'does this decision conform to architectural intent?'

The accumulated gap between the architectural decisions a system contains and the architectural decisions any human on the team actually understands — a structural human limitation that scales with every AI coding tool you adopt.

The accumulated set of design decisions that knowingly deviated from Steering intent — approved under time pressure, resource constraint, or scope trade-offs — recorded with a named owner, stated justification, and known risk. Distinguished from accidental drift by the presence of deliberate choice.

The principle that every CDA decision is fully traceable — which lenses ran and why, which did not fire and why not, what the AI saw, what file or line triggered each verdict. Explainability is what lets developers interrogate verdicts and AI agents self-correct autonomously via the rkito session protocol.

Determinism is not a natural property of LLMs. rkito achieves deterministic governance outcomes through procedurally coded knowledge for everything that can be a rule, and the narrowest possible AI context for everything that cannot. An open engineering and research problem at the core of reliable AI governance.