Decoding AI's Open-Source Course Maps Three Ways to Run an Agent Loop and the Provider Economics Behind Each

In LangChain's Terminal-Bench experiment, changing only the harness—keeping the same model—moved a coding agent from roughly 30th place into the top 5. The article covers Paul Iusztin's open-source course, published through Decoding AI, which builds Decode, a Python agent defined in ~20 lines using Pydantic AI. Decode separates three run modes—interactive, remote, and async—each with different latency profiles and inference providers. It notes Claude Code's core loop is roughly 150 lines, with everything else being harness.