In my previous post, I explained why I dropped GitHub Copilot for a custom AI workflow using Roo Code and OpenRouter. I needed more control, better context, and no rate limits.
Now, it’s time to stop talking and start building.
In this episode, we are going to set up the entire environment, configure our multi-agent team, and most importantly, learn the difference between “Vibe Coding” and actual Agentic Coding.
Here is how I set up my “Dev Team in a Box” to build a complex Business Central integration.
Step 1: The Engine Room (OpenRouter)
The first step is to get our “fuel.” I use OpenRouter because it allows me to access top-tier models (like Claude 4.5 Sonnet and Gemini 3 Pro) on a pay-as-you-go basis.
- Create an account: Go to openrouter.ai and sign up.
- Add Credits: You only pay for what you use. I usually add about $20, which goes a long way.
- Get Your Key: Generate an API key. You will need this to connect Roo Code to the brain of the operation.
Step 2: Configuring the “Agents”
Once Roo Code is installed in VS Code, we need to assign roles. This is where the magic happens. We don’t just use one model for everything; we assign the best model for each specific job.
- The Orchestrator: I use Gemini here. The Orchestrator acts as the Project Manager. It needs a massive context window to keep the entire project scope in its “head,” and Gemini excels at this.
- The Coder: I switch to Claude 4.5 Sonnet for the actual coding agent. It is incredibly fast and produces high-quality syntax.
- The Debugger: Also powered by Claude.
This setup ensures that when the Orchestrator hands off a task, the specialized agent executes it with the best possible tool.
Step 3: Establishing the Rules (The .roo Folder)
AI is only as good as the instructions you give it. If you want consistent code, you need Rules.
I structure my project with a specific .roo folder containing markdown (.md) files that define my coding standards. These aren’t just prompts; they are persistent guidelines the AI must follow.
For this project, I added rules for:
- Naming Conventions: e.g., Object names must not exceed 30 characters.
- Design Patterns: Enforcing the Single Responsibility Principle.
- Testing: Mandating that all code must be testable and include unit tests.
By saving these as .md files, I create a reusable template. I don’t have to repeat myself every time I start a new project.
Step 4: The Spec (Planning Before Coding)
This is the biggest mistake people make with AI: they just start chatting and coding. That is “Vibe Coding.” We want Agentic Coding.
Before a single line of AL code is written, I ask the Orchestrator to create an implementation.md file.
I feed it the requirements:
- We need a WebShop integration using the Staging pattern.
- Data validation errors shouldn’t block data delivery.
- We need API pages, List pages, and specific Codeunits for processing.
The AI then breaks this down into a phased implementation plan. This allows me to review the architecture before we burn tokens on code generation. If the plan is wrong, the code will be wrong.
Step 5: Initializing Memory
Finally, I initialize the project’s Memory Bank.
I use a command to tell Roo Code to create a memory structure. This creates files that track:
- Project goals
- Active context
- Progress status
This is crucial. When I shut down my computer and come back tomorrow, the AI “remembers” exactly where we left off, what has been completed, and what is next.
Ready to Build
We have our engine (OpenRouter), our team (Roo Code Agents), our laws (Rules), and our blueprint (The Spec).
In the next episode, we will finally let the agents loose to start writing the code.
All code can be found on github: dfredborg/AIDevelopmentWithoutFilters