AI Expands Software Boundaries But Deterministic Systems Persist

Software Spectrum and Classification

The author defines software broadly: system software (OS, databases, compilers), large applications (Office), middleware (Android, Kubernetes), front-end apps, business ERP, microservices, industrial software, and AI itself. These categories differ fundamentally. A key distinction is between software requiring high determinism (errors cause severe issues) and error-tolerant assistive software (customer service, chat, video generation).

Two Core Perspectives

First, AI expands the software spectrum, offering new solutions for problems traditional software handled poorly — this is AI's core value. Second, the author rejects the view that AI will rewrite, consume, or eliminate existing deterministic software. There is no viable business case for discarding valuable, efficient, battle-tested domain knowledge accumulated over long periods.

Five Methods for Solving Problems with Computers

Developers or domain experts build LLM-based agents directly.

Domain experts use ready-made agent platforms (e.g., OpenClaw) and write skills or tools.

Developers use code agents to write deterministic software, then review and test before delivery.

Traditional deterministic software development.

Domain experts use binary agents (binary programs) as proposed by Musk.

Methods 1, 2, 3, and 5 are new capabilities from large-model advances that greatly expand the problem-solving boundary. The core capability of LLMs is accepting multimodal input (natural language, images), understanding, planning, reasoning, and generating outputs with inherent uncertainty (code, plans, tool calls).

Business Logic as the Deciding Factor

Any execution method must close the business logic loop. The choice depends on the specific problem and whether a lower-cost, better method creates competitive advantage. Rewriting working deterministic systems lacks a clear business case.

Why Code Agents Generate Deterministic Code

Code agents are currently the most successful LLM application. They generate deterministic software rather than executing tasks directly because: (a) the output requires extremely high determinism and human accountability for quality; (b) per-invocation token cost is high, yielding low ROI. Generating programming-language-expressed software is the optimal compromise, implicitly acknowledging the long-term value of deterministic software.

Traditional Software Will Persist

Most traditional software will remain because long accumulation has made processes highly efficient and automated, with determinism guarantees. The business logic for rewrite — what real user value does it bring? — is hard to articulate. The "hammer-nail" fallacy applies. Experts suggesting AI handles rapid business changes and customization (e.g., data analysis) where deterministic software was inefficient may have a point, but current industry focus stays on problems deterministic software couldn't solve well: deep research, video generation, code generation — scenarios tolerating lower determinism or requiring human review.

Case Studies: 12306 and Salesforce

The 12306 ticketing system illustrates that replacing a button click with a voice command adds little value; the complex business logic, design decisions, and constraints evolved over time are underestimated by replacement advocates. A cited expert points to Salesforce as a potential example of agents directly handling deterministic business — a case to watch.

The Challenge of Quality Attributes in AI-Generated Code

Software engineers value reliability, performance, quality, security, compliance, observability, maintainability. As hand-written code decreases, system thinking becomes more critical. The core question: how to ensure AI-generated code possesses these quality attributes? Code is the final software artifact; how can AI understand and construct a structure with these properties? Will AI automatically create suitable architectures? Current industry exploration uses specifications (specs), but specs are simplifications — overly detailed specs become code themselves. Can AI automatically bridge the gap?

Conclusion

AI technology evolves, but foundational logic remains: what are human end-needs and real problems? The article emphasizes analytical process over hype, grounding each claim in business rationality and engineering pragmatism.