For developers, the pressure is always mounting to build code more and more quickly and efficiently. AI has been improving drastically year over year mainly due to increasing training data corpus and increasing compute — and AI-powered tools like GitHub Copilot have revolutionized how we work by offering rapid, in-line suggestions that turbocharge workflows and boost productivity off the charts. 

Researchers and companies are exploring new ways to improve AI by getting multiple AI systems, or “agents,” to work together toward a shared goal. This approach, called Multi-Agent systems, means combining different AI models so that they can each contribute based on their unique strengths. And it’s been shown to be much more effective for complex tasks than using just one model.

In software development. research has already shown that multi-agent AI systems can produce good results. Academic research (like the ChatDev chat-powered software development framework) and some companies (like Devon) have tested this idea. 

Up until now, it hasn’t been common to see “multi-agent” AI setups being widely used in commercial products for software development. But all that is changing. 

In part, that’s because speed isn’t the only goal of a developer — especially if that acceleration is speeding the work in the wrong direction. A suggestion that is misaligned with requirements can actually set the project back. 

At Crowdbotics, we’re deeply familiar with the difference a well-crafted product requirements document (“PRD”) makes. By establishing a solid PRD, business stakeholders’ needs can be communicated clearly and precisely to developers and AI alike, setting up a more stable launch pad for the entire project. Our Crowdbotics PRD-AI requirements generation tooI is designed around that principle, helping non-developers communicate their needs in a way that developers can quickly understand and take action. 

So we know how much creating a good PRD improves exponentially through the use of AI. And we know that AI coding assistants are most valuable when they’re grounded in a project’s requirements.

Enter the multi-agent model. 

So we wondered, given that GitHub Copilot is proving to be invaluable to developers, was there untapped potential for connecting Copilot’s suggestions directly with PRD insights — helping it to be not only more powerful, but also more accurate? In other words, could we give GitHub Copilot more business context that it could leverage to ensure the code it suggest to developers is more accurate? Is it possible to have two, independent AI tools work together to improve overall performance without additional model training or increased computation? 

The answer, as you’ll see, is a very definitive yes.

Crowdbotics’ GitHub Copilot Extension

We built the Crowdbotics GitHub Copilot Extension tool to bring together the power of AI-driven coding assistance with the clarity of AI-generated requirements context. By facilitating the creation of a complete PRD and then feeding it directly into the GitHub Copilot environment, we aimed to improve Copilot’s suggestion relevance. Feeding Copilot with specific business and project context, we believed, would help it generate more accurate, targeted suggestions and streamline the coding process.

Proving the Tool’s Effectiveness

To prove this hypothesis — and measure the impact of this approach, we designed an experiment, which we recently documented in a detailed white paper. Our goal was to validate whether embedding requirements context would significantly improve GitHub Copilot’s code suggestions. (Spoiler: they do. By a lot.)

To conduct the tests, we divided 101 developers into three groups, each tasked with completing the same coding assignment, structured as follows:

1. Control Group: Using a standard code editor without GitHub Copilot.
2. Copilot Group: Using GitHub Copilot without any additional PRD context.
3. Copilot Enhanced Group: Using GitHub Copilot seeding with business requirements from Crowdbotics PRD-AI tool.

Participants were given a challenging back-end project, designed to require a substantial understanding of business requirements and coding precision. We measured two main metrics: the rate at which developers accepted Copilot’s code suggestions and the rate of task completion. 

The results were clear:

• PRD Alignment Increased Suggestion Acceptance: Developers in the Copilot Enhanced Group were 13.8% more likely to accept Copilot’s suggestions compared to the standard Copilot Group, raising the baseline acceptance rate from 27% to nearly 41%. This represented a 51% improvement, underscoring how context makes suggestions more relevant and actionable.
  
• PRD Alignment Resulted in Higher Task Completion Rates: The enhanced setup resulted in a 24.5% higher completion rate compared to the Control Group and a 14.8% higher completion rate compared to the standard Copilot Group. These improvements confirm that integrating business context enhances Copilot’s overall utility.

For those interested in the technical details, our full white paper explores the experiment methodology, data points, and outcomes in depth. Read the full white paper here.

Explore the Crowdbotics GitHub Copilot Extension

Our experiment highlights how AI tools can be refined to create smarter, more efficient development cycles. The Crowdbotics GitHub Copilot Extension represents a significant step forward in AI-enabled development by demonstrating that PRD-driven alignment can lead to faster, more accurate coding.

We invite you to explore the new Crowdbotics GitHub Copilot Extension and see how intelligent AI alignment can revolutionize your development process. Contact us to learn how Crowdbotics can help accelerate your development cycle with AI and CodeOps.

Interested to see the Crowdbotics GitHub Copilot Extension in action?
Download it from the GitHub Marketplace today.