Live Demo: Building a Custom AI Research Team with Task Decomposition On Claude Sub-Agents

In this episode of The Build, Cameron Rohn and Tom Spencer demonstrate a live build of a custom AI research team using Claude Sub-Agents and task decomposition to drive practical development decisions. They begin by walking through a hands-on walkthrough by example, showing how Claude Task Decomposition spawns specialized research sub-agents that coordinate via API integration and memory systems connected to a Local Chroma Instance. The conversation then shifts to tooling and deployment: they compare Langsmith for observability, Vercel for front-end hosting, Supabase for lightweight persistence, and MCP tools for orchestration and developer workflows. They explore technical architecture decisions such as the Sub-Agent Complexity Cap, agent-to-agent communication patterns, and strategies for embedding memory and local search to reduce latency. Along the way they address building in public strategies, discussing transparent iteration, open-source AI toolkit models, and monetization pathways for niche research assistants. The episode emphasizes developer ergonomics with concrete examples of agent code, error handling, and CI/CD considerations. In closing, they advise entrepreneurs and developers to adopt modular sub-agent architectures, iterate publicly with clear telemetry via Langsmith, and prioritize composable APIs—a forward-looking call to build iteratively and ship pragmatic AI products.