Most programming systems draw a hard boundary between the reasoning layer and the control layer, requiring a validation step at every interface and forcing programmers to maintain two models of their system simultaneously. We present \emph{Relational Agenda Programming} (RAP), a uniform execution model in which reasoning, control flow, and I/O all flow through the same relational machinery with no privileged escape hatches. The key architectural choice is a flat-stack virtual machine that executes goals via an explicit work queue, coupled with a pending-query agenda that is itself a first-class relational term. Queries receive the current agenda as input, reason over it using unification — the same machinery used for domain reasoning — and return a \emph{ChangeSet} specifying atomic agenda modifications. ChangeSets are valid by construction: the $no-ops$ and $cons-ops$ constructors are relations that fail on invalid operations, so invalid ChangeSets cannot be expressed. We present a C++20 implementation with zero dependencies, embeddable via FFI from any language. We validate the model through two case studies running against a working implementation: embedded security policy verification demonstrating the Probe meta-evaluation primitive, and reflective agenda reasoning demonstrating a query that inspects and modifies pending work through the same relational machinery used for domain reasoning.