My path to my first coded Rovo agent

Like many Agile coaches, I was excited when I first got access to Rovo, Atlassian’s AI tool. With just a prompt, it can surface information, refine stories, summarize project history, and even automate simple tasks through custom agents.

But I wanted to go further. I wanted to automate more complex workflows so I could spend more time focusing on the real needs of my teams. That meant building a coded Rovo agent. I wasn’t starting from zero — I had some experience with Python scripts and small automation projects — but Rovo requires JavaScript or TypeScript, two languages I had never used. That’s when I turned to ChatGPT as my tutor.

Learning the foundations

My goal was clear: create an agent capable of generating a full performance analysis for a team — highlighting key metrics, trends, and progress on epics. The first step was understanding how coded agents actually work. I used AI to teach me the basics: when to use them, what happens behind the scenes, and what Forge is — the platform that powers Rovo apps and agents.

Once I had that foundation, I asked ChatGPT to help me build the simplest coded agent possible. I didn’t want a ready-made answer; I wanted a guided learning experience. So I prompted it to act as a tutor and create a step-by-step mini-course. This helped me install the dependencies, understand the agent manifest, and create and deploy a simple “hello world” agent.

Coding the agent

Then came the real challenge: coding the agent I actually needed, in a language I barely knew.

From previous experience, I felt like starting with natural language was the easiest way to translate ideas into code. So that’s what I did. I wrote out how the agent should be triggered, what information it would need (variables), and the actions (functions) it would have to perform to complete the analysis.

This approach worked far better than simply asking ChatGPT to “write an agent that does XYZ.” For simple agents, that might be enough. But for anything more complex, it’s incredibly easy to forget specific scenarios or exceptions the code must handle. That leads to a long back-and-forth of “add this,” “consider that,” or “if X is true, do NOT do that.” It can turn into a messy process.

I wish I could say that my agent worked in the first try, but that would be far from the truth. My terminal showed several errors the first time I ran it.

At first, I simply copied the errors into ChatGPT, but I quickly realized it would suggest a possible fix, not necessarily the best one. What worked far better was pasting the error and prompting my tutor to “help me troubleshoot this so I can understand the root cause.” That led to several iterations of reading and adding logs, analyzing output, and eventually finding the root causes.

After a few rounds of problem-solving, my code worked. I then prompted my tutor to review my code the way a senior developer would — refactoring it, removing the logs and helper functions I no longer needed, and making sure it followed good software development practices. After a clean-up and another test, the agent did exactly what I needed.

A few lessons

If you’re thinking about building your first coded Rovo agent, here’s my two cents, based on my experience:

• Start with understanding, not copy-paste: Let AI guide you through the concepts behind the code.
• Use natural language as your starting point: Describe what you want the agent to do — that will help you get a clear idea of what you need.
• Build the simplest version first: A basic “hello world” agent teaches you more than you’d expect.
• Troubleshoot: Don’t rush to the final result. Each step builds your understanding.

Final Thoughts

Everyone’s journey will be different. I already had some exposure to software development, and that helped — but what made the real difference was taking the time to understand the concepts behind Rovo agents and using the tools available to learn them properly. If you take the same approach, you’ll set yourself up for success, no matter your starting point.