Martin Fowler’s Guidance on AI‑Driven Software Development

As AI coding tools become deeply embedded in enterprise development, teams experience faster code delivery but also see declining system maintainability, frequent production issues, and accelerating technical debt. The article argues that these problems stem not from the tools’ capabilities but from traditional software engineering’s inability to accommodate the non‑deterministic collaboration mode introduced by large language models.

01. Core Essence: AI‑Induced Non‑Determinism

Conventional software engineering relies on deterministic execution—compilation, test runs, and function outputs are reproducible and traceable. Large language models are probabilistic; the same instruction can yield different implementation paths, produce unverified explanations, and cause unexpected side‑effects, injecting large‑scale non‑determinism into the development pipeline.

02. Scenario Boundary: Two Coding Modes

Vibe Coding leverages AI for rapid prototyping, one‑off scripts, and low‑value tools, lowering the barrier for non‑developers but failing to support long‑term asset growth because it interrupts the software‑engineering learning loop.

Agentic Engineering complements Vibe Coding by emphasizing code ownership, system understandability, and long‑term maintainability. It introduces disciplined engineering practices that keep the team in control of the system.

03. Engineering Foundations: Value of Traditional Practices

AI‑generated code does not diminish the importance of testing, refactoring, CI/CD, and static analysis; instead, these practices become even more critical. Small, granular commits constrain model divergence, reduce review costs, and improve fault localisation and rollback efficiency.

The recommended workflow adopts a two‑layer collaboration: the large language model handles intent understanding and solution exploration, while deterministic tools (compilers, IDE refactoring, static analysers) execute and verify the code.

04. Adaptation Core: Harness Architecture

The Harness acts as the adaptation layer linking AI’s non‑determinism with deterministic engineering systems. An Agent is defined as Model + Harness. The Harness comprises modules such as file system, sandbox, state management, tool invocation, hook rules, verification mechanisms, and long‑task control, forming the backbone for engineering‑grade AI deployment.

Consensus from OpenAI, Martin Fowler, and industry experts yields three Harness principles: version‑manage complex task plans and decision logs; enforce architectural rules via linters and CI rather than static documents; continuously iterate small releases to clear technical debt; translate model errors into system constraints.

05. Risk Control: External Verification

AI‑generated output can masquerade errors with fluent language, even fabricating test results. Therefore, safety must rely on quantifiable external feedback. Teams should build multi‑dimensional validation covering test execution, type checking, dependency boundary detection, operation approvals, and production monitoring.

Permission management, a core Harness component, should be tiered by risk: low‑risk actions auto‑execute, medium‑risk require human confirmation, high‑risk demand strict approval and audit, while untrusted inputs and private data are isolated to prevent injection attacks.

06. Role Upgrade: Engineers as Supervisors

With AI handling repetitive coding, engineers transition from direct coding to supervisory engineering: defining tasks, organising context, overseeing model execution, evaluating outputs, and codifying error rules. Control points shift to planning, boundary setting, permission governance, and system evolution, sharpening the focus on architectural value.

Although code generation costs drop, tasks such as requirement analysis, architecture design, risk validation, and security governance become more critical, raising the overall complexity of software engineering.

07. Practical Measures: Six Foundations

Lightweight task decomposition – prioritize independent, verifiable tasks like test completion, bounded bug fixes, or call‑chain analysis to avoid large‑scale refactoring and lower review/rollback costs.

Knowledge‑base consolidation – convert design decisions, architectural constraints, and historical issues into ADRs and rule files that the model can read as standardized context.

Pre‑built verification system – strengthen core‑path tests, boundary checks, lint rules, and dependency controls to provide rapid feedback for AI automation.

Permission risk stratification – classify operations (file I/O, test runs, dependency changes, DB access, external calls) by risk level and apply differentiated approval policies.

Error standardisation – break errors into categories (parameter exception, environment failure, permission denial, timeout, test failure) for precise localisation.

Experience systematization – after fixing model‑generated errors, transform solutions into test cases, lint rules, and task templates to continuously enrich the Harness.

08. Conclusion

AI continues to accelerate code generation, but core engineering responsibilities remain unchanged. Non‑deterministic collaboration demands a robust deterministic governance framework. The goal is not full automation but to encode engineering experience as system rules, allowing AI to operate safely within controlled boundaries and achieve a balance between efficiency and quality.