Agentic & Vibe Coding: The Future of Development? 🚀

After a few days of training on generative AI, I’ve spent the past few weeks exploring what lies behind two recent concepts: vibe coding and agentic coding. So here’s a little feedback from my experience on the topic. 😊

🔍 First things first: What do these two terms mean?

Vibe coding refers to software development assisted by Large Language Models (LLMs) such as ChatGPT, Copilot, and others. Through iterative conversations and prompts, development happens progressively, guided by intuition and experimentation. 🎨

Agentic coding, on the other hand, involves developing with AI agents that interpret a specification, break it down into tasks, orchestrate them, and interact with the system (filesystem, browser, terminal...). There is very little human intervention, since agents can be authorized to read logs, launch applications, execute commands, open a browser, and so on. 🤖

✨ Vibe Coding

💻 My context:

PyCharm, ChatGPT (free version), building a small Flask web application in Python.

⚙️ How it works:

Despite the limitations of a free version of ChatGPT (yep, don’t expect hours of coding without paying 😅), with patience and testing after each iteration, the app gradually takes shape. A basic cycle of prompt → code → test → adjustments becomes the core workflow. 🔁

✅ The result:

After a few days, you get a documented codebase that smoothly integrates session-based authentication, CSS, JavaScript, and even functional documentation — all that without having to write too much code manually. ✨

🤖 Agentic Coding

💻 My context:

VS Code with the Roo Code extension, using the mistral/devstral-small:free model and the SPARC methodology (Specification, Pseudocode, Architecture, Refinement, and Completion), once again to build a containerized Flask web app.

⚙️ How it works:

This time, you write a much more comprehensive prompt — typically specifying the agent’s role, the rules it must follow for each SPARC phase, and of course the project’s goal, functional requirements, use cases, technical and UI/UX constraints, and the testing methodology (TDD). 📝

The SPARC agent then starts working more or less autonomously depending on the permissions granted, switching between agents as needed — acting in turn as architect, coder, tester, etc. The agent announces the actions it will take, specifying which agent role it assumes at a given time. Checkpoints are created regularly, making it easy to rollback when necessary.

✅ The result:

Despite the limitations of using a free model (slow responses, some tasks not completing), I still ended up with a tested, documented application, ready to be deployed as a container. 📦

🤔 Open Questions

My opinion is based on two projects built from scratch — but what about integration into an existing project? I intend to conduct tests on it shortly.

I only used free models, and I can only imagine how powerful the paid models must be in terms of performance and code quality — especially Claude 4… right Jérôme Revillard ? Coming soon 😊

✅ Verdict

Two rooms, two vibes in the age of generative AI! 🎭

After a few weeks of experimentation, my first impression is that vibe coding is more suited to prototyping and building POCs, boosting the productivity of experienced and creative developers. Agentic coding, on the other hand, seems more focused on long-term productivity, code quality, and suggests a new way to rethink the developer’s role. 🛠️📐

One or the other? Maybe both. Either way, one thing’s for sure: this is the future of software development, and we’d better not miss the turn! 🔀🚀