AI Agent Harness: Critical for Performance & Cost

Many teams focus on the AI model when building agents, but new findings point to the "harness" as more important. In one experiment, changing only the harness moved a coding agent from 30th to the top 5, using the same model throughout. This means how you run the agent loop is a critical architectural decision. A new open-source course, "Building a Coding Agent From Scratch," introduces a Python agent called Decode. It demonstrates three distinct run modes for an agent, each with a different latency profile and suited for a different inference provider. The core of the system is a headless harness, meaning it has no user interface. This harness runs the standard agent loop where the AI model picks an action, a tool executes, and the observation feeds back. The agent itself is small, with most of the complexity residing in the harness. The three modes are: interactive online, remote offline, and async online. Each mode has specific characteristics, like how input is handled or whether a human is actively monitoring. This highlights that the choice of how you run an agent significantly impacts its performance and cost.