OpenAPI/AI

Inspiration

The initial API design phase often suffers from lengthy negotiations between developers and clients. We believe AI can bridge this gap. While Large Language Models (LLMs) may struggle with direct code generation, they excel at human interaction and context retrieval. By leveraging this strength alongside established standards like OpenAPI, we can streamline API development significantly.

What it does

Our multi-agent system acts as an Intelligent API Architect. It begins with a Product Manager Agent that intuitively interacts with the user, asking clarifying questions to define API requirements. Once all necessary information is captured, this agent passes a structured specification to a "development team" of specialized agents. This team collaboratively generates the OpenAPI (Swagger) definition, produces server stubs and client SDKs, and creates comprehensive documentation. Finally, the system initiates a pull request to your GitLab project, seamlessly integrating the new API into your existing development workflow.

How we built it

We leveraged Google Cloud's Agent Development Kit (ADK) to construct a robust multi-agent architecture. This system utilizes diverse patterns including sequential execution for critical handoffs (like requirements gathering to specification), parallel agents for concurrent code and documentation generation, and generator/validator agents to ensure quality outputs. To interact with the outside world, our agents employ tools that interface directly with OpenAPI Generator for code production and GitLab APIs for version control. Every step, from conception to commit, is intelligently managed by our agents and their specialized tools.

Challenges we ran into

A primary challenge was the inherent limitation of LLMs like Gemini in directly producing complex, perfectly reliable code. To overcome this, we made a strategic design choice: while LLMs handle intelligent interpretation and orchestration, deterministic, industry-standard OpenAPI code generation libraries perform the actual code output. This hybrid approach ensures both intelligent design and reliable, production-ready code.

Accomplishments that we're proud of

We are particularly proud of our robust multi-agent architecture. By meticulously defining each agent's role and choreographing their interactions (sequential, parallel, and validator patterns), we created a highly modular and efficient system where every component contributes significantly to a smooth, automated workflow.

What we learned

This project provided invaluable hands-on experience with ADK and the intricacies of designing a multi-agent architecture. We gained practical knowledge on integrating intelligent agents with external services like GitLab and leveraging the broader Google Cloud Platform ecosystem.

What's next for OpenAPI/AI

Due to time constraints, the final deployment phase wasn't fully completed. Our immediate next step is to extend the system to automatically deploy the generated API stubs to Google Cloud Endpoints and Cloud Run, completing the entire end-to-end development and deployment pipeline for an intelligent API lifecycle.

Built With

• adk
• cloud-run
• gemini
• gitlab
• openapi