Agent engineering: how to design a mobile application in 3 weeks with AI agents

The potential of generative artificial intelligence (GenAI) fuels all strategic discussions, but how does it translate concretely on an industrial scale? At Smile, we prioritize proof by practice over mere announcements.

During a recent webinar, the Neopixl team (our mobile application development expert), represented by David Renoux (Digital Experience Director at Smile and CEO of Neopixl), Déborah Delin (Digital Consultant), and Frédéric Bovy (Engineering Manager), shared their experience using their agentic factory . Their challenge: to design and deploy a functional iOS application in just three weeks, presented at the VivaTech trade show using our client Speed Vegas as a case study.

Here are the main takeaways from this session, including methodological breaks, the evolution of our tech jobs and realities on the ground.

AI maturity levels: from "vibe coding" to industrialization

To structure the integration of AI throughout the entire development cycle (specifications, UI/UX design, code, testing, security), Neopixl relies on three levels of technical maturity:

• Vibe coding: ideal for the rapid exploration phase, creating Proof of Concepts (PoCs), or trade show demonstrations. The goal is absolute speed, often at the expense of managing technical debt or architectural sustainability.
• Assisted coding: the developer remains at the center of the IDE (development environment) and collaborates with code copilots (like Tabnine , used by the Smile teams). While effective at accelerating the pace, the human element remains the main bottleneck in the process.
• Spec-driven development: This is the disruptive methodology chosen for this 3-week project. Here, the source of truth changes sides: it is no longer located in the final codebase, but upstream, within the design, functional specifications and extremely structured acceptance criteria (using tools like GitHub's SpecKit ).

The expert's perspective: AI is a power multiplier. If the input data (the backlog or design tokens) is unclear, the AI agent will fill in the gaps by hallucinating, which propagates cascading errors. Rigorous planning from the outset is the essential condition for industrial efficiency.

Case study: The premium mobile app "Speed Vegas" for VivaTech

In order to validate the robustness of this agentic factory, the team took on the challenge of building from scratch a premium user experience for Speed Vegas customers (supercar driving sessions on the track).

Beyond the classic booking screens, we wanted to integrate complex technological features: a facial scanning module to assess the size of the pilot's helmet and an augmented reality (AR) interface for real-time customization of their equipment.

Lessons learned and the reality on the ground

In open source engineering as in AI, transparency is part of our DNA. Deborah shared a textbook case encountered at the beginning of development: an onboarding screen that generated with an unexpected neon green color and completely ignored the expected interaction mechanic ( swipe to start ).

Analysis of the error revealed that a color was missing from the Figma design system and that the onboarding logic was not sufficiently explicit in the specification file (in Markdown format). The development agent therefore applied a default value. This pitfall demonstrates the crucial importance of human review at the start of a project. By immediately correcting the functional specifications and the design system, the error was permanently eliminated for the rest of the production chain.

The impact of AI on tech jobs: augmented roles

Does the rise of agent-based factories replace tech experts? Our teams' answer is unequivocal: absolutely not . Human expertise is valued more highly, with the focus shifting towards higher value-added tasks.

• The software architect: their role is expanding. In addition to designing the target application architecture, they orchestrate and configure the agentic production chain (choice of models, configuration of agents, definition of AI developer/reviewer roles, integration of mobile security standards, accessibility, automatic synchronization of tasks to Jira).
• The developer: they gain a broader perspective. Freed from writing repetitive code tasks (less than 1% of the code on this application was written manually), they act as a quality assurance supervisor. When faced with a bug, they don't hard-code it, but update the machine's skill to permanently fix the issue.
• The product owner (PO): by delegating technical documentation and the laborious creation of tickets to AI, the PO can refocus on their true value: the consistency of the product vision, functional creativity and the final user experience.

Governance and security: protecting enterprise code and data

Integrating artificial intelligence agents into development workflows requires a strict governance framework to eliminate any risk of intellectual property leaks. For this project, Neopixl implemented isolated environments and specific API configurations (via Anthropic) prohibiting the use of our data or source code for model training.

For large-scale industrial projects or those subject to stringent regulatory constraints, the Smile Group designs architectures based on sovereign LLM models , hosted on trusted cloud infrastructures (such as Outscale or Scaleway). This closed-loop approach guarantees the absolute security of your digital assets.

Conclusion: Agent engineering is "production ready"

The results of this three-week sprint are conclusive: AI-driven engineering is production-ready. It offers unprecedented speed, drastically reduces the regression rate across iterations, and puts innovation back at the heart of delivery cycles.

Next steps: register for our webinar series!

The AI agent revolution isn't limited to mobile. To explore how these technologies are transforming the entire digital ecosystem, join the experts at Smile in upcoming episodes of our series:

• Episode 2: Next-generation CMS: When agentic AI reinvents the Web and Drupal
  In-depth analysis of the impact of AI agent architectures on the design of complex web platforms and the evolution of content management ecosystems.
  Register for the Web & CMS webinar
• Episode 3: Moving to agentic reality: the complete interconnection of customer experience with ERP
  Discover the inner workings of our AI orchestrator and learn how to automate data and communication flows between your applications, your e-commerce platform (Shopify), your data architectures (Snowflake) and your ERP (Odoo).
  Register for the Interconnection & ERP webinar