From 'Vibe Coding' to Agentic Software Engineering: Why We Are Open-Sourcing Our Quality Gates Framework
"Building software with artificial intelligence agents has become astonishingly fast. Yet without strict governance, generated code swiftly drifts toward technical sloppiness and hidden debt. At DevSupAi, we are open-sourcing the Agentic Engineering System (AES) to enforce uncompromising software engineering discipline on AI agents."
Standard engineering edition: includes deep architecture layers, technical terms, and deterministic Quality Gate specifications.
AI velocity is meaningless without deterministic guardrails: the Agentic Engineering System (AES) structures agentic workflows into 5 strict layers, enforces epistemic honesty (MEASURED vs ESTIMATED), and guarantees SEO, accessibility, and security compliance before any production deployment.
1. The pitfalls of 'Vibe Coding': three critical hazards of ungoverned AI
Developing software applications using autonomous AI agents such as Claude Code, Cursor, Codex, or Antigravity makes assembling components and interfaces possible in mere minutes. This is the era of Vibe Coding.
Under the hood, however, early euphoria often yields to a brutal reality: ungoverned AI-generated code inexorably drifts toward what the industry terms software slop (accumulated technical debt, redundant dependencies, and silent regressions).
At DevSupAi, we strictly reject delivering sloppy code to our clients. Inspecting unconstrained AI workflows highlights three systematic failure modes:
- •Context contamination (Prompt Bloat) :Piling all directives into a monolithic instruction file (CLAUDE.md, .cursorrules) leads the model to conflate universal principles, stack rules, and obsolete client data, triggering hallucinations of pricing or internal business logic.
- •Empty promises and commercial hallucinations :By default, large language models aim to please the user. They spontaneously fabricate fake testimonials, unverifiable marketing percentages, or assert that a user interface is 100% WCAG AA compliant without testing any screen reader.
- •Complacent delivery sign-offs :The agent triumphantly announces that everything is production-ready without having executed any real unit tests, without auditing responsive viewports, and without measuring bundle payloads.
2. Our answer: the 5-Layer Context Separation Rule
This structured separation ensures that the AI agent receives the exact relevant context required for each specific task, eliminating attention dilution and preventing cross-layer pollution.
The 5-layer architecture of the Agentic Engineering System (AES) framework
| Layer | Scope | Operational role | Concrete DevSupAi implementation |
|---|---|---|---|
| Layer 1: GLOBAL | Universal engineering principles | Non-negotiable overarching directives: security, non-regression, zero fake reviews, zero emojis in production | Absolute prohibition of fabricating pricing or unverified percentages |
| Layer 2: PROCEDURES (Skills) | Deterministic automation scripts | Locally executable tools producing tangible, quantified verification checks | Automated A11y checks, sitemap.xml validation, exact bundle size accounting |
| Layer 3: PROJECT | Single declarative configuration | Verified source of truth (project.config.json) defining legal identity, domain, locales, and official pricing | Official service packages (€950, €1,850, €2,600, €3,200, €400 daily rate) |
| Layer 4: STACK | Targeted technology constraints | Framework and runtime specific rules without polluting other software stacks | React 19 standards, vite-react-ssg hydration protocols, SSR isolation |
| Layer 5: FACTS | Authentic business domain facts | Human-validated data; strict ban on extrapolating or guessing missing information | Official legal identity of Alexandre Pabst EI, contact details, and authentic case studies |
3. Epistemic Honesty: eliminating arithmetic and marketing guesswork
Large language models excel at conceptual reasoning, but they are notoriously unreliable when performing mental arithmetic or physical benchmarking. In a performance summary, an unconstrained agent tends to invent bundle sizes or round percentages arbitrarily.
Whenever a derivable value exists (such as the sum of HTML + CSS + JS payloads or an average calculation), the AI agent is forbidden from computing it mentally: it must invoke a dedicated script and replicate the exact output.
The 4 mandatory epistemic statuses in our Quality Gate governance
| Status | Rigorous definition | Validity criteria | Application example |
|---|---|---|---|
| MEASURED | Data obtained from a locally executed script or automated test | Instrumented proof required | Exact JavaScript bundle payload = 89.17 kB gzipped |
| NOT MEASURED | Metric impossible to simulate locally or not yet instrumented | Mandatory explicit labeling if real-world data is absent | Real-world Core Web Vitals (LCP, INP, CLS) without active RUM field monitoring |
| ESTIMATED | Projective calculation or theoretical model | Mandatory disclosure of computation methodology | Theoretical bandwidth savings projected prior to staging release |
| QUALITATIVE | Technical assessment stemming from automated or human code review | Never equates to a physical measurement | Qualitative observation regarding component decoupling |
4. Automated SEO, AEO, and accessibility safeguards by default
Technical SEO and digital accessibility do not tolerate approximations. Our Quality Gates orchestrator automatically verifies critical engineering criteria prior to release:
- •Heading hierarchy :Exactly one <h1> element per page, followed by a strictly sequential structure (<h1> to <h2> to <h3>) without level skipping.
- •Canonical URLs and internationalization :Automated generation of absolute canonical URLs and reciprocal hreflang annotations across every language variant.
- •llms.txt standard for AI search engines (AEO) :Strict markdown link formatting enabling conversational engines (SearchGPT, Perplexity, Claude) to index and cite verified business facts without hallucinating.
- •Honest accessibility testing :Prohibition of claiming 100% accessible or WCAG compliant based solely on automated linters; reports must itemize tested criteria versus those requiring manual audits.
5. Why we are releasing the Agentic Engineering System as Open Source
Artificial intelligence is fundamentally reshaping software development. We believe that the true value of an engineering studio like DevSupAi lies not in raw keystroke speed, but in the rigor of its guardrails, the resilience of its architecture, and the standard of excellence delivered to each client.
By releasing the Agentic Engineering System (AES) under the permissive MIT license, we fulfill two central missions:
- •Sharing a reusable industrial standard :Providing developers and engineering teams with an operational foundation compatible with Claude Code, Cursor, Codex, Antigravity, Aider, and Copilot.
- •Guaranteeing full transparency for our clients :Demonstrating that every showcase website, e-commerce shop, or custom web application built at DevSupAi rests on proven, transparent, and auditable software foundations.
Explore the GitHub repository and collaborate
The project is available today under the open-source MIT license on GitHub: https://github.com/devsupai/agentic-engineering-system.
You are welcome to clone the repository, test our Quality Gate scripts, and adapt the 5-layer rule to your own development pipelines.
Looking to engineer a modern web application at AI velocity backed by uncompromising engineering rigor? Get in touch with DevSupAi to discuss your roadmap.
Too technical? You can switch to our plain English edition without jargon at any time.
Frequently Asked Questions
Recommended Resources & Related Packages
Alexandre Pabst
Independent web designer and developer, founder of DevSupAi in Saint-Mihiel (Meuse, France). Custom showcase websites, e-commerce stores, and tailored web applications.
Have a project in mind?
Get a free technical feasibility study and a personalized quote within 24 hours.