Architecture as Program: Capability-Injected, Model-Driven Software Development in the Age of AI Agents

AI coding agents have removed the thirty-year obstacle to graph-based code representations: the human preference for text. This proposal describes a development model in which the primary artifact is a signal graph (a functional reactive program with explicitly typed capability boundaries) that serves simultaneously as architecture model, security policy, and source of truth. Implementation code is typically a compiled artifact generated by AI agents within capability-restricted sandboxes; existing code can be incorporated as opaque nodes with explicit capability signatures, giving a migration path that does not demand rewriting up front. We synthesise functional reactive programming, object-capability security, and trust-annotated information flow into a single architectural substrate, and argue that this synthesis is newly practical due to the convergence of agent-authored code, lightweight sandboxing (WASM/WASI, BEAM), and capability hardware (CHERI). The proposal presents a three-phase research agenda, from core demonstrator through formal verification, and identifies the open problems that must be resolved to deliver on its claims.