LIVE AI Coding Challenge - YOU Choose What I Build (Using Kiro!)

Key Concepts

• Audience-Driven AI Development: Building an application based on audience votes, fostering engagement and a unique learning experience.
• Iterative Development & System Evolution: Employing a “Plan-Implement-Validate” (PIV) loop and prioritizing addressing root causes of errors over simple fixes.
• AI Agent Workflow: Utilizing AI agents powered by LLMs (Claude Opus 4.5, GPT-4) to automate coding tasks, guided by steering documents and validated through browser automation.
• Obsidian Integration: The project focuses on building a voice agent to interact with a personal Obsidian knowledge base.
• Tool Prioritization & Simplification: Starting with the simplest possible implementation (read-only operations) and prioritizing efficient tools over complex pipelines (e.g., favoring markdown search over RAG for code).

Project Initiation & Setup (Part 1)

The stream began as an audience-driven AI coding challenge, leveraging a voting application built with Anthropic Harness to determine the project. Viewers submitted 37 ideas, casting 86 votes, with the project leaning towards an Obsidian voice agent. The development workflow centers around a PIV loop, utilizing tools like Claude Opus 4.5 (accessed via a $20/month Kira subscription), AquaVoice ($10/month), Verscell Agent Browser, and potentially Pyantic AI & LiveKit. The project is partially framed within the Kira-hosted Dynamis Hero Hackathon, offering a $17,000 prize pool with an extended deadline of January 30th. A quick start template within Kira was used to initiate the project, generating initial steering documents (product.md, structure.md, tech.md).

Refining Scope & Building the Foundation (Part 2)

The initial plan to use MCP was abandoned in favor of a read-only tool operation scope, focusing on search, summarization, and detail extraction from Obsidian documents. The importance of iterative development and updating steering documents to reflect the revised scope was emphasized. A “shared tools layer” was identified as the fundamental first step, preceding agent and front-end development. Validation was prioritized, starting with simpler components. The developer explored Convex as a potential database replacement for Superbase and highlighted the Verscell Agent Browser’s suitability for AI agent automation compared to Playwright. The developer advocated for tool-based exploration over indexing codebases for RAG, aligning with Anthropic’s approach.

Voice Agent Development & Initial Testing (Part 3)

The focus shifted to building the voice agent component, integrating LiveKit and utilizing the Versell agent browser for planning. The agent successfully generated a plan to connect LiveKit to the front end and update environment variables. The decision was made to avoid a full RAG pipeline for the codebase, opting for a simpler markdown and blob search. An unexpected issue with missing API endpoints was resolved by identifying a missing REST.py file, demonstrating the need for careful verification. The agent successfully generated a comprehensive plan, including end-to-end testing and a readme file, though initial testing failed due to unstarted services.

System Evolution & Completion (Part 4)

The final segment emphasized “system evolution” – addressing the root causes of errors rather than just fixing individual instances. A bug requiring cleanup of started services was resolved by updating the “execute” command in the steering documents to automatically stop services after testing. This iterative process of identifying issues, brainstorming solutions with the agent, and implementing changes was highlighted as the core of the workflow. The Dynamis Agent Coding Course was repeatedly mentioned as a resource for deeper understanding. The stream concluded with the codebase published on GitHub, inviting community contributions. Challenges with GPU discovery and slow response times were addressed by switching to the OpenAI Realtime Mini model.

In conclusion, the stream successfully demonstrated a practical, audience-driven approach to AI-assisted software development. The emphasis on iterative development, system evolution, and careful tool selection provided valuable insights into building reliable and effective AI agent workflows. The project, an Obsidian voice agent, served as a concrete example of these principles, showcasing the potential of AI to automate and accelerate the coding process while highlighting the continued importance of human oversight and validation.