The Ten Core Principles of Software Engineering 3.0

In the era of large‑model and autonomous agents, the author proposes Software Engineering 3.0 (SE 3.0) as a paradigm shift that goes beyond the Agile manifesto’s 12 principles and the so‑called SE 2.0. SE 3.0 is built on four value pillars articulated in the SE 3.0 Declaration and is operationalized through ten interlocking principles.

Intent‑First, Acceptance‑Criteria‑Driven Development

The starting point of SE 3.0 is a crystal‑clear business intent expressed as executable acceptance criteria (AC) in Gherkin format. This transforms vague requirements into low‑entropy instructions, reducing information uncertainty (information‑theoretic view) and providing precise error signals for rapid feedback (control‑theoretic view). The AC becomes the immutable “constitution” of the development system, guiding both Builder Agents that generate code and Verification Agents that validate it.

Human‑AI Symbiosis

Human experts define high‑level “why” and “what” – values, ethics, strategic trade‑offs – while AI agents handle the “how”, executing at scale, pattern‑recognizing, and iterating quickly. This creates a “dumbbell” team structure: humans at the ends (product, architecture, QA) and AI in the middle automating code generation, testing, and optimization, with a Human‑in‑the‑loop/ Human‑out‑of‑the‑loop hybrid for critical decisions.

Data‑First and Continuous Knowledge Accumulation

Business and development process data become the foundation of LLM capability and organizational intelligence. High‑quality, structured private data outweighs model size, as research shows. Knowledge is distilled through four layers: raw data (L1), structured knowledge graphs (L2), experience patterns such as prompt libraries (L3), and automated feedback loops that evolve the knowledge base (L4), forming an “external brain” for the organization.

Minimalist Process, Maximal Feedback Density

SE 3.0 replaces heavyweight process with ultra‑light feedback loops that are fully automated by AI agents. Feedback speed, precision, and reliability replace detailed process documentation. CI/CD pipelines evolve into Agentic DevOps where test generation, code review, and defect fixing are performed by agents, freeing humans to focus on intent definition and final arbitration.

Full‑Lifecycle Model Integration

Large models are no longer mere coding assistants; they are integrated across the entire lifecycle—from intent capture, design, code generation, testing, to operations. Continuous context flow between agents eliminates information silos and reduces re‑work, with platform engineering providing the necessary infrastructure for seamless data hand‑off.

Adversarial Quality Assurance

Quality assurance relies on heterogeneous agents (e.g., Claude 4.6 as Builder, Gemini Pro 3.1 as Verifier) to create a genuine adversarial dynamic, preventing the “mode collapse” seen when generator and discriminator share the same model. This ensures robust detection of systematic biases and defects.

Executable Acceptance Tests as Living Documentation

Acceptance tests serve simultaneously as human‑readable specifications and machine‑executable validation artifacts. Their executability guarantees that documentation stays in sync with code, while self‑maintenance and traceability link every test back to its originating business intent.

Model Asset Management

Models that generate code become core assets requiring version control, performance monitoring, security audits, explainability analysis, and continuous fine‑tuning—essentially applying MLOps practices to keep model capabilities aligned with evolving business needs.

Complex Adaptive System (CAS) Perspective

SE 3.0 is described as a continuously self‑adapting system where each failure yields a refined “bug pattern” that enriches a quality gene pool. Micro‑, meso‑, and macro‑level knowledge flows (builder‑breaker feedback, project‑level knowledge graphs, organization‑wide knowledge sharing) create emergent order, making the software factory increasingly resilient and “anti‑fragile”.

In conclusion, the ten principles are presented as a belief system that guides teams through unprecedented technological change, aiming to build an intelligent, self‑evolving software factory where intent drives execution and AI amplifies human value.