One persistent misconception about AI in software development is that it replaces developers. It doesn’t. AI takes over the routine tasks, but decisions regarding architecture, quality, and security remain the work of humans. The developer’s role is shifting from writing code to providing direction. We’ll show you how we use AI in the development process, where it works, and where it doesn’t.
-min.jpg&output=webp&q=90&hash=MTczODE2MjczNQ==&w=800&h=800&fit=cover&a=center)
AI-powered software development is a process in which experienced engineers purposefully use AI to build software faster and with higher quality. AI is integrated throughout the entire process: from code and testing to exploring architecture and detecting bugs. In areas where judgment is critical, a senior developer remains in charge. The result is a shorter turnaround time without compromising the quality of the final product.
This is what it’s all about for us. The question isn’t whether you use AI in development, but where you draw the line between what a model does and what a human decides. If you draw that line incorrectly, you’ll end up with code that looks convincing in the demo but behaves incorrectly in production. That error doesn’t always show up right away, and that’s what makes it costly. As AI takes on more operational tasks, that line becomes more important, not less. That’s why senior review isn’t an afterthought for us—it’s an integral part of the process.
The principles behind our approach
AI provides suggestions; an experienced developer determines whether they are correct. Architectural decisions, exceptional cases, and the assessment of whether something should be automated at all remain the work of humans. It is precisely in those areas where “erroneous but plausible” errors are hardest to spot that a senior developer steps in. This prevents a convincing-looking error from going unnoticed until it reaches production.
Working faster is pointless if quality suffers as a result. AI adheres to the conventions we define in advance, and linting (automatic code checking), testing, and review collectively catch what a model itself cannot detect. That doesn’t slow us down; it makes our speed reliable.
We aren’t tied to a single coding agent, and you retain ownership of your code and data. Governance can be reviewed retrospectively, so you can always see what has happened to your information. This way, the software remains yours, even if the technology changes or you decide to go in a different direction later on.
No single tool can cover all of this. GitLab is the hub where everything comes together: context, code management, CI/CD, and governance—all in one place. We integrate coding agents like Claude Code with it, so we’re not tied to a single agent but can choose the best fit for each situation. Right now, that’s often Claude, which excels at in-depth reasoning during refactoring across multiple files and targeted code reviews, but we’ll switch to a different agent as soon as that proves to be a better fit. Within GitLab itself, we also use the Duo Agent Platform for agents running there. This way, you can work with the best agent for the task at hand, without being locked into today’s choice.
The fact that GitLab is at the heart of it all is no coincidence. Cube has been an official GitLab Channel Partner since 2024 and has been using GitLab since 2018. We were also mentioned in the announcement in which GitLab expanded the Claude integration into the Duo Agent Platform, precisely because of the combination of capable AI and robust governance. The fact that we’re mentioned there speaks to how early we’ve been putting this approach into practice. When it comes to tools, we deliberately remain flexible so we can adapt to what works. You can read about how that governance is embedded in the pipelines on the page about DevSecOps with GitLab.
Tools for Every Stage
AI vult code aan, genereert boilerplate (de standaard, terugkerende coderegels die elk project nodig heeft) en refactort over meerdere bestanden. Dat haalt het repetitieve werk weg bij de engineers en levert code op die consistent de afgesproken conventies volgt, bewaakt door linting en review.
AI generates test cases, keeps existing tests up to date, and flags regressions—that is, functions that stop working after a change. This increases test coverage without the manual work causing the schedule to slip. A human determines which risks and edge cases really matter.
AI compares multiple architectural approaches, identifies trade-offs, and tests assumptions. This speeds up the planning phase and provides a stronger basis for the choice. The final decision rests with an architect or senior developer. AI broadens the perspective; a human makes the final call.
AI analyzes error messages, logs, and code to identify the likely cause of a bug, especially for errors deep within a system. When combined with error tracking tools like Sentry, you can pinpoint the root cause more quickly rather than just the symptom. The developer reviews and confirms the solution, as false positives do occur.
An agent addresses an issue within defined parameters, modifies code across multiple files, runs tests, and prepares a merge request. The engineer is the director, not a typist. What the agent is allowed to do, when it stops, and where human approval is required are all defined in advance. At Cube, this is now handled by Claude Code, using the GitLab Duo Agent Platform to provide context and governance in one place. We’re not locked into this solution: if another coding agent performs better for the task, we’ll switch.
Taken together, these steps result in software that goes live faster, has fewer errors in production, and remains maintainable as you continue to build. The day-to-day work gets done faster, while you retain control over your code, your data, and the direction.
A question that often comes up in initial discussions: Does our code go to external AI services? That only happens as a deliberate choice, not by default. The cloud tools run on enterprise versions with IP protection configured, and it’s determined in advance what may and may not be sent to external AI services. For strict requirements, self-hosting is an option. The underlying governance is based on role-based access (permissions per role) and an audit trail for each AI action via MCP, so you can see what happened afterward. Do you want your data to be ready for that? It starts with a Data Management Platform.
Together, we’ll figure out what really needs to be done. Which processes are causing bottlenecks, which tasks AI can take over, and which ones will intentionally remain in human hands.
This translates into a concept and an architecture. AI explores the possibilities here, while a senior developer sets the course.
Iterative development using AI-assisted coding, automated testing, and agentic engineering within defined frameworks. Each release is tested and reviewed.
Once the system goes live, we’ll monitor it, make improvements, and continue developing it, with error tracking and clear agreements on management.
Nearly 15 years of building software in a constantly changing market—that’s what determines where we do and don’t trust AI. That experience is the difference between accepting an AI suggestion and knowing when to set it aside. It also determines how we adapt to an organization’s needs. Some clients want Cube to build the software. Others have their own team that wants to learn how to work with AI tools and are looking for a partner to guide them through the process. We can do both.
Do you want to use AI without losing control? We’re happy to help.
OpenAI is an advanced AI technology that can conduct conversations, answer complex questions and make information understandable. It helps organisations make knowledge accessible to employees and customers, turn data into actionable insights and handle routine queries automatically.
Claude (Anthropic) can process information and unlock knowledge without exposing or using sensitive data to train the AI model externally. This allows organisations to apply AI securely within internal tools, portals and business processes.
n8n is an open-source workflow automation platform that allows you to easily connect different applications and data sources. It allows you to automate tasks and processes, synchronise data and build workflows that fit exactly the way your organisation works.
Gemini is an advanced AI platform that combines natural language and other data sources to automate and improve business processes. It can interpret information, perform analysis and perform tasks within different applications and workflows.
Laravel, a leading PHP framework, accelerates development with tools like Eloquent ORM, Blade templating, and built-in security features. It provides an efficient and reliable foundation for building web applications.
LangChain is a framework that enables language models to be linked to data sources, applications and workflows. It enables organisations to automate complex tasks, enrich knowledge bases and internal tools with AI, and make customer portals smarter.
PHP, or Hypertext Preprocessor, serves as a powerful scripting language for creating dynamic web pages and applications. It allows you to develop robust and flexible online solutions tailored specifically to your needs.
Python is known for its efficient performance and speed, making it ideal for data analysis, machine learning, and web development. With a wide range of modules and an active community, developers can create dynamic and customized solutions.
Vue.js, an innovative JavaScript framework, enables you to create dynamic and interactive web applications. It offers not only speed and flexibility but also a modular architecture suitable for both simple and complex projects. Its excellent documentation and performance make it even more appealing.
Let's connect
Talk to engineers who use AI in their development process every day. We’d love to exchange ideas about where this field is headed.
With courage, a level-headed approach, and a focus on quality.
Amber Projectmanager
_lagekwaliteit.-2.jpg&output=webp&q=90&hash=MTc1NDI5MTI2NA==&w=540&h=660&fit=cover&a=top)
Bart Developer
Bernard Developer
Bob Test Engineer | Security & Privacy
Demi Officemanager
Dennis Full Stack Developer
Guus Full Stack Developer
Jarno Business Director
Jarno Back-end Developer
Jasper Developer
Jenne Digital Marketeer
Jeroen Back-end Developer
Job CEO | Executive
Joost Tech Lead | Team Pentagon
Jordy Software Architect
Justin Tech Lead | Team Hexagon
Kees Projectmanager
Kevin Developer
Laura Digital Designer
Maartje Content Marketeer
Mans Operations Manager
Marcus Projectmanager
Marleen Projectmanager | Team Lead Octagon
Mart Front-end Developer
Melanie Projectmanager | Team Lead Hexagon
Mick Tech Lead | Team Nexus
Nicky Projectmanager
Onno Business Controller
Remi Software Architect
Rogier Business Consultant
Rosan Manager People & Organization
Roy Full Stack Developer
Ruben Full Stack Developer
Ruben Digital Designer
Sander Developer
Stijn Front-end Developer
Tamara Business Consultant
Thomas Back-end Developer
Tom Tech Lead | Team Octagon
.jpg&output=webp&q=90&hash=MTc3MTg0MzIwMw==&w=540&h=660&fit=cover&a=top)
Wesley Projectmanager
No. AI is taking over an increasing share of the hands-on work, but decisions regarding architecture, quality, and security remain the responsibility of experienced engineers. As a result, the developer’s role is shifting from writing code to providing direction. As AI becomes more capable, that guiding role becomes more important, not less.
Yes. Cube develops the software or provides guidance to your own team as they learn to work with AI tools. The GitLab Duo Agent Platform brings together tools, agents, and development in a single environment, with the oversight you need. This allows you to maintain control in-house while gaining experience with this way of working.
No. The GitLab Duo Agent Platform is at the heart of our development environment, but we aren't tied to a single coding agent. We choose what works best for each situation and adapt when something better comes along, without vendor lock-in.
Only as a conscious choice, not as the default. The cloud tools run in enterprise versions with IP protection, and it is determined in advance what data may be sent to external AI. If requirements are strict, self-hosted is an option.
Software architecture is the structural blueprint of an application: the decisions about how components are divided, how they communicate, and how the system handles growth and maintainability. These are the early, fundamental choices that guide everything that follows.
With MCP, the integration layer runs on your own infrastructure. You control which data is visible per tool and who has access. MCP supports authentication via OAuth or API keys, rate limiting, and logging. Data does not leave your environment unless you explicitly allow it.
