Skip Agent Frameworks at First: Build One Small Agent Loop by Hand

A developer who spent months studying agent frameworks and buying hardware still could not build a working agent, which the article presents as a sequencing mistake rather than a lack of effort. The author argues that builders should start with a minimal stack of just a model and a loop, avoiding frameworks like LangChain and CrewAI until the underlying problems are felt. Writing the loop by hand makes concepts such as state, retries, and errors concrete, and it makes later abstractions like LangGraph easier to understand. The piece notes that local runners with OpenAI-compatible endpoints, such as Ollama and llama.cpp, make this approach easy to start with modest hardware. It suggests beginning with a single small agent, such as a research assistant with a few tools and a step cap, then adding memory or tests only where they measurably help.