The Third Restructuring of Software Form: From the... | AI Research

The Third Restructuring of Software Form: From the Three-Tier Architecture to Storage, Models, and Agents proposes that software architecture is evolving from a three-tier model into a system composed of three core elements: a generalized database, a large model, and an agent. The authors, Wei Lin, Tao Zhou, Zhaofei Xie, and Changgui Hong, argue that as the cost of using models to reason and generate behavior decreases, traditional software layers will be absorbed or elevated, leaving only storage, intelligence, and execution as the essential components of software.

The Convergence of Three Elements

The authors define the terminal form of software as a composition of three specific roles:

• Generalized Database: A unified abstraction for all persistent state, memory, and knowledge. It serves as the system's "past," providing auditability and transactionality.

• Large Model: The "present" of the system, acting as the intelligence core responsible for reasoning, decision-making, and generating interfaces or actions.

• Agent: The "future" of the system, functioning as an execution loop that connects the model to the database. It manages the plan, memory, and tool-use cycle.

Collapse of the Three-Tier Architecture

The paper argues that the traditional three-tier architecture—user interface, business logic, and data—is being restructured. The user interface layer is expected to dissolve into on-demand generation by models, though the authors note that critical information must still be presented through deterministic projections of stored data.
Business logic is re-partitioned into three categories: 1. Model Reasoning: Expressible, non-critical rules handled by the model. 2. Storage Constraints: Critical, declaratively expressible rules (such as uniqueness or foreign keys) embedded directly into the database. 3. Deterministic Tools: Critical, non-declarative logic (such as complex optimization solvers) that remains as verified code.
The data layer is elevated to the sole persistent infrastructure, as it provides the necessary foundation for the model to maintain long-term state and ensure system reliability.

Evidence from Prototypes

To test the role of the database as an arbiter of correctness, the authors implemented a prototype for an intelligent production-scheduling system. Using SQLite, they enforced machine exclusion and precedence rules as database constraints. Even when the "planner" (a simulated model) introduced errors into its proposals, the database rejected all violations, maintaining a feasible schedule. This suggests that the system can guarantee correctness through storage constraints, regardless of the quality of the model's reasoning.

Boundaries and Limitations

The authors emphasize that this thesis is not a universal solution. They identify specific conditions under which the approach holds: the task domain must be expressible, verifiable, externally stateful, and tool-complete. The thesis is limited by factors including:

• Determinism: Some logic cannot be fully absorbed by models or declarative constraints.

• Cost: The overhead of agentic loops and model calls must be considered.

• Security and Verifiability: The authors acknowledge that models can hallucinate, which is why they argue that critical logic must be anchored in deterministic storage constraints rather than relying solely on the model's output.