A preprint examining AI-centered software found a sharp split between a model that proposed schedules directly and one that could call a solver. In 20 direct Qwen-Plus trials, no schedule met all the reported constraints, while storage rejected every violating schedule. In 10 solver-delegation trials, the system reported the correct schedule length in all 10 and stated that the result was feasible in all 10.
Three jobs, three components
The paper’s proposal gives three jobs to three components: storage keeps the application’s state, a large model handles flexible reasoning, and an agent carries out actions and uses tools. In the authors’ formal description, the traditional software form is reduced to storage, model and agent when the task domain permits it. Business logic is divided by expressibility and criticality: some rules stay with model reasoning, some become storage constraints—database rules that refuse invalid entries—and some are sent to verified deterministic tools.
The thesis is conditional, applying only where tasks are expressible, verifiable, externally stateful and tool-complete. It does not cover strongly deterministic, formally verified core tasks; there, the paper says logic remains in storage constraints or verified code.
A database as a gatekeeper
To make the idea concrete, the authors built a Python/SQLite production-scheduling prototype with 10 machines and 200 unit-time operations. Machine exclusion was enforced with a UNIQUE(machine, slot) rule, while a BEFORE INSERT trigger checked precedence—the required order between operations.
The storage-arbiter test used planner-error rates of 0, 0.1, 0.2 and 0.3, with 200 proposals at each level. It recorded which proposals were rejected, what persisted and whether the resulting schedule was feasible. At 0.1, 0.2 and 0.3, storage rejected 16, 33 and 45 violating proposals, respectively, and the persisted schedule remained feasible even when up to 30% of assignments were perturbed.
A separate test deliberately omitted the capacity check from hand-written code. The table marked that capacity violation as “leaked”, while the storage implementation was marked “caught”.
The hard calculation still went to a solver
A second experiment used OR-Tools CP-SAT on five reproducible synthetic job-shop instances. All listed instances were marked optimal, but solve times ranged from 0.01 to 9.67 seconds and reported schedule lengths ranged from 594 to 1,143.
The live model’s weak point
The live-model test ran Qwen-Plus through a production API, asking it to schedule a job-shop instance directly without a solver tool in 20 independent trials. The table reported an apparent mean schedule length of 8.9 against an optimum of 9, with mean latency of 3.84 seconds. Machine-exclusion violations occurred in all 20 trials and precedence violations in 18; storage rejected all 20 violating schedules.
With solver delegation, the solver was requested in all 10 trials. The system reported the correct schedule length—9—in all 10, stated feasibility in all 10, and recorded a mean latency of 4.82 seconds.
A conditional argument, not a replacement claim
The paper reports descriptive counts and timings rather than inferential uncertainty. Its evidence consists of a conceptual architecture and small synthetic scheduling demonstrations, with no conventional end-to-end baseline; the broader applicability conditions are argued conceptually rather than validated across diverse real-world systems.
The reported tests therefore support a narrower point: in this scheduling setup, storage constraints rejected invalid proposals and solver delegation supplied correct feasibility and schedule-length reports. The broader thesis remains limited to expressible, verifiable, externally stateful and tool-complete domains, while strongly deterministic, formally verified core logic remains in storage constraints or verified code.
Paper data and sources
Original title: The Third Restructuring of Software Form: From the Three-Tier Architecture to Storage, Models, and Agents
Authors: Wei Lin, Tao Zhou, Zhaofei Xie, Changgui Hong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text