Redesigning Software as Storage Models and Agents

Summary & Methodology Analysis

The paper formalizes the transition to Software 3.0 by defining a system as a composition of a generalized database, a large model, and an agent. The collapse mechanism partitions business logic based on expressibility and criticality. Logic that is highly critical but difficult to express declaratively is offloaded to tools like OR-Tools CP-SAT, while other tasks are handled by the model. The reference architecture uses a stateless intelligence core, a heterogeneous persistent storage layer, and an agent execution loop to process requests. In a prototype using Python and SQLite with 10 machines and 200 operations, the storage layer acted as a gatekeeper for the model output. When tested with Qwen-Plus, the storage layer successfully identified and rejected all twenty infeasible scheduling proposals, ensuring the persisted state remained valid. The model relies on probabilistic reasoning, which is a technique where the system predicts likely outputs based on statistical patterns rather than strict logical rules. The paper identifies critical boundaries for this architecture. It is not suitable for systems where errors are unacceptable, such as aerospace applications or medical dosing, as the model cannot provide the necessary formal guarantees. Furthermore, the architecture requires a feedback loop that relies on objective verification. If a task result cannot be verified, the agent cannot detect model hallucinations, which are instances where the model generates confident but false or nonsensical information. Finally, the paper notes that for high-frequency or extreme low-latency scenarios like trade matching or real-time recommendations, freezing logic into dedicated, non-model-based implementations remains the only viable choice.