Summary
Cole Medin tackles a question increasingly relevant to AI developers in 2025 and beyond: when should you build agents on top of batteries-included SDKs like the Claude Agent SDK or Codex SDK, and when should you still reach for traditional frameworks like Pydantic AI, LangChain, or N8N? The video argues that both approaches remain valid — but for very different use cases — and lays out a clear decision framework.
Medin identifies three core limitations of SDK-based agents: significantly higher latency due to reasoning overhead, greater token consumption (and the associated cost implications), and reduced determinism since the SDK manages much of the agent’s behavior. For production deployments requiring sub-second response times, full observability over message history, or multi-user scale beyond a personal subscription, he recommends building from the ground up with a framework. Conversely, for personal productivity tools, internal automations, and rapid prototyping, the SDKs offer a compelling shortcut. He also covers the evolving role of RAG in modern agents, noting a shift toward skills-based tool definitions and MCP servers over traditional chunking-and-retrieval pipelines.
Code examples are provided in a linked GitHub repo, and Medin briefly previews a four-hour workshop on building a ‘second brain’ with the Claude Agent SDK scheduled for April 3rd in his Dynamis community. This video is a practical reference for developers navigating the increasingly fragmented agent tooling landscape.
📺 Source: Cole Medin · Published March 26, 2026
🏷️ Format: Deep Dive
