Onboard

Inspiration

Software documentation is often the most neglected part of the development lifecycle, yet it is the most critical for scaling a team. The inspiration for this project came from the realization that while code is structured and logical, the way humans learn is through storytelling. We wanted to move beyond static, dry Markdown files and build a system that treats a GitHub repository like a textbook, capable of teaching itself to a newcomer.

What it does

Onboard is a multi-agent platform that dynamically converts complex software repositories into interactive, narrative-driven documentation. It utilizes a specialized suite of agents to analyze repository structures, craft educational narrative chapters, and generate corresponding quizzes to assess a user's understanding of the code.

How we built it

The system is built on a Supervisor-Worker architecture using Amazon Bedrock AgentCore and the Strands SDK.

Supervisor Orchestrator: A FastAPI backend running Claude Sonnet plans documentation routes and triggers sub-agents via the JSON-RPC Agent-to-Agent (A2A) protocol.

Specialized Sub-Agents: We deployed three distinct A2A workers: a Code Explorer to fetch GitHub source code via Lambda, a Narrative Writer to craft educational content, and an Assessment Creator for dynamic quizzes.

Frontend: The user interface is a React SPA built with TypeScript and Vite for submitting targets and reading generated output.

Challenges we ran into

One of the most significant hurdles was the cloud deployment of the distributed sub-agents to the Bedrock AgentCore Runtime. We had to manage the iterative process of using the agentcore CLI to push specialist agents and manually sync their generated Agent ARNs into our environment configurations. Additionally, implementing the A2A protocol was difficult because it required maintaining a stateless environment where the Supervisor had to manage the entire conversation context perfectly, as the workers lack their own persistent memory.

Accomplishments that we're proud of

We successfully orchestrated a distributed system where specialized AI agents operate collaboratively rather than in isolation. A key achievement was the ability to translate raw GitHub repository trees into a structured, narrative curriculum that is both human-readable and pedagogically sound.

What we learned

We gained deep insights into LLM orchestration and the efficiency of distributing complex cognitive tasks across multiple specialist agents. By delegating specific roles—such as code retrieval, narrative crafting, and quiz generation—to separate entities, we found that the overall quality and accuracy of the documentation improved significantly. We also refined our skills in deploying cloud-native AI infrastructure using the Amazon Bedrock AgentCore ecosystem and managing complex environment configurations with AWS CDK.

What's next for Onboard

The next phase involves expanding the Code Explorer Agent's capabilities to support a wider range of version control platforms beyond GitHub. We also plan to enhance the Narrative Writer to support multi-language output, allowing international teams to generate localized curriculum-style documentation automatically.

Built With

• agentcore
• amazon-web-services
• github-mcp
• python
• react
• strands