AI agents are already transforming how software is built. They can dramatically accelerate certain tasks, but they can also get stuck on details that would seem obvious to a human. They are also changing how people learn, collaborate, and organize work within tech teams.
So, is it still worth learning to code? Le Wagon Canada brought together three guests to compare the promises of AI with the realities they encounter in practice. The discussion was moderated by Jean-François Berthiaume, a Le Wagon Canada alumnus.
Our guests:
Editor’s note: The discussion has been condensed and grouped by theme to make it easier to read while remaining faithful to the views expressed.
Francis Lacoste: To me, AI is like a jetpack. It lets you move incredibly fast and fly, but it can also send you straight into a wall if you’re not careful.
Eric Cappannelli: I use the image of railway tracks under construction. We’re building the foundations of something that will eventually allow us to move much faster, without yet knowing exactly how. For an SME, the risk is either believing the train has already arrived or, at the other extreme, dismissing AI as just another shiny new toy.
Antoine Ayoub: I would describe it as quicksand. The ground is beginning to stabilize, but a tool released today could be obsolete tomorrow. That is concerning, but also exciting: we need to relearn certain things and rethink how we work.
Eric Cappannelli: The experience alternates between major “wow” moments and moments of real frustration. On some projects, the system can prioritize tasks, track progress, and consult the code. It’s a bit like having a junior developer beside me whose work I review.
But AI can also spend hours stuck on simple details. When I tried to produce an audio track with a convincing Quebec accent, I spent more time adjusting the wording and intonation than I would have spent having an artist read it. More recently, while working on a moderately complex project involving several technologies, I decided to start again from a blank page rather than continue patching a foundation that was heading in the wrong direction. At that point, AI had cost me more time than it had saved.
On the other hand, I have almost entirely automated my timesheets, their synchronization with my clients’ tools, and part of the billing process. I can now spend that time delivering more value to my clients.
Francis Lacoste: Some experienced developers tried AI last year, got poor results, and held on to that first impression.
The conversation changes when they experience what some people call a “Claude Code moment.” A worthwhile project that no one had previously made time for is suddenly completed in two hours, with surprisingly good quality. That’s when you think, “OK, this is a new world.”
Francis Lacoste: AI usage still varies widely within the same team. Some people manage several agents, others mainly use AI for autocomplete, while others want nothing to do with these tools yet.
Three models are starting to emerge. In the “Iron Man” model, one person uses several agents independently. In the shared model, each person works separately, but the team pools its methods and resources. Finally, in the still-uncommon “multiplayer” model, several people collaborate with agents on the same project.
Knowledge management therefore becomes essential: how can individual discoveries be shared so the entire team can progress?
Eric Cappannelli: We also need to talk about the exhaustion caused by overusing agents. You can have an idea in the morning and put it into production that afternoon. The machine has no limits, but we do.
When several tasks are moving forward at the same time, we are constantly multitasking and switching contexts. Some people have already hit a ceiling in terms of fatigue. We therefore need to maintain personal discipline and healthy working habits.
Another skill is becoming mandatory: delegation. When we get a good result from AI, it is because we have successfully delegated the task to it.
You need to know what you are doing and then learn how to have it done. With an agent, you need to explain what to do, anticipate certain problems, and specify how to proceed and verify the result. It is like working with an intern or a junior team member. This presents a genuine human challenge and requires a new kind of learning.
Francis Lacoste: Training the next generation of senior developers is a real challenge, and there is no good answer yet.
Experience helps develop a “smell test”: the intuition that tells you when a project is heading in the wrong direction. You can develop it on your own, but that takes time. Mentorship and apprenticeship can accelerate the process.
One option would be to bring back pair programming with an LLM, or large language model. A junior team member could drive the tool while a senior points out the areas to watch. This model remains uncommon, however.
Antoine Ayoub: LLMs can support learning, but we need to avoid letting them replace the work our brains need to do. Excessive or inappropriate use can weaken that muscle. In an intensive training program, certain breakthroughs are essential to developing a solid grasp of the fundamentals.
It is a bit like taking a Formula 1 car to buy a baguette. If you ask an LLM to draft a contract in an area you do not understand, it can produce something highly intelligible but extremely complex. That abundance of information becomes exhausting when you no longer know how to navigate the response. The same thing happens in training: a simple question can receive a complicated answer, even though the course had deliberately made the concept accessible.
I also remain convinced of the value of human support. When you are discovering a new field, it is difficult to ask about something you do not yet know or to build a coherent training program on your own. I made more progress in a few months with my ukulele teacher than I did trying to learn by myself for several years.
Later in the discussion, Francis returns to critical thinking.
Francis Lacoste: You develop it through practice, by reviewing your work and having someone validate your critique. LLMs can contribute to that process: one agent produces the code while others examine it from different angles. But if you do not practise critical thinking yourself, you cannot delegate it.
Francis Lacoste: LLMs allow us to automate tasks that we can define clearly or complete them in collaboration with AI. This also applies to product and design specialists, who can quickly create prototypes through vibe coding (the practice of describing in natural language what you want to build). These prototypes can be used to validate an idea before being handed to a development team, which can rebuild them in a robust and reliable way.
Eric Cappannelli: Vibe coding reminds me of what Excel or Access enabled in an earlier era. People who are close to the business can translate their knowledge and expertise into software.
Antoine Ayoub: What people have learned throughout their careers is becoming even more important. We can combine our previous experience with new foundational technical skills. The models can then help us go further.
Francis Lacoste: I do not think AI fundamentally changes the qualities that make someone a good developer. We are still looking for curiosity, the ability to learn and solve problems, an understanding of the domain, and empathy for users. In a professional context, the value does not come from the code in isolation, but from the problem it solves. I appreciate beautiful code and elegant architectures, but if they do not solve a business problem, their value remains limited.
At the end of the panel, a participant asks whether technical tests are still relevant in job interviews.
Francis Lacoste: To me, the result of an exercise matters less than a candidate’s ability to discuss what they did, explain their reasoning, and demonstrate their ability to learn. A library that is important today may change in six months or two years. General knowledge and transferable skills therefore remain the most important things to assess in an interview.
Eric Cappannelli: Yes, but not for much longer. It is a bit like getting a driver’s licence without becoming a mechanic. I think today’s Wild West will eventually become industrialized and that we are not far from having more autonomous coding systems. We can still debate when that will happen.
Antoine Ayoub: Yes. Without foundational knowledge, you cannot develop genuine critical thinking about what an LLM produces. Copying and pasting answers in a field you do not understand quickly becomes tedious and unsatisfying.
Francis Lacoste: Yes, just as we need to learn to write, though not necessarily to become writers. We need to learn to code so we can understand, read, and think, while recognizing that “the job won’t be coding.”
Would you like to develop the foundational skills required to build with AI and evaluate what it produces? Explore Le Wagon Canada’s bootcamps.
At Le Wagon, we believe that “Code is here to build products”. Our Tokyo campus
In this Q&A session, we dive into the inspiring career transformation of Sofia, a Le
After completing a part-time Web Development bootcamp, Ines launched her freelance career and embraced a
Mark and Marisa joined us here in Lisbon this past June to learn how to
Axel is one of our Data Analytics alumni, originally from France and now based in
Le Wagon is a space for career switchers, aspiring entrepreneurs, and even
entrepreneurs looking
