It is the harness, not the model, that carries the agent: the layer everyone conflates
The software layer around the model carries the system prompt, tools, the agentic loop and your freedom to swap models. When an agent misbehaves, the culprit is usually the harness, not the model.
Anyone who works with agents knows the moment: the system does something absurd, and hours go into asking whether the model is broken or your own scaffolding is. An essay published by Earendil hands that argument its missing vocabulary by defining the layer everyone conflates with the model: the harness. Flagged by developer Jeremy Morgan, the piece keeps the definition plain: an agent harness is a piece of software that provides an environment for an AI model to operate within. The analogy comes from climbing: a harness holds the climber, organizes their equipment and lets them move safely. The climber still climbs, but nothing works without the harness.
What a harness is made of
System prompt: instructions injected into every conversation that govern the model's behavior without being embedded into the model itself. The essay's comparison is apt: like the instructions a new employee gets on their first day, layered on top of what training already taught.
Tools: capabilities written in code, from web search to code execution to composing email. The crucial distinction: the harness makes tools available but does not dictate their use. The model decides which tool to call and when.
Agentic loops: the cycle where the model assesses results and decides whether one more pass is needed. The essay's example is a school comparison: the model searches, judges whether the data suffices, builds a spreadsheet, and loops back if something is missing. Retries and verification passes live in this loop.
Translation layer: the abstraction that lets one harness work with models from different providers. By preventing vendor lock-in, it keeps leverage in the user's hands.
Why it gets conflated, and why that matters
Agents are often summarized as "model + harness", and that shorthand hides the harness's own architectural role. When debugging, the distinction is vital: context assembly, tool routing, retries and verification loops are the harness's job. If an agent reaches for the wrong tool or repeats the same mistake, the fault usually lies not in the model's intelligence but in the software wrapped around it.
The essay's deeper argument is about ownership: unlike proprietary AI applications, you can own, modify and run a harness locally. That takes power and leverage away from the AI labs and puts it in the hands of end users, giving you the freedom to compare models and keep your own correspondence history. It cites Pi, a minimal open-source harness with more than 5,000 community extensions, the email-based Lefos, and the chat-based OpenClaw.
This week we covered Hermes, where agents DM each other, WebMCP, where sites open up to agents, and OpenClaw, which runs on your own devices. Their common denominator is exactly this layer: it is becoming clear the race runs as much on the harness around the model as on the model itself.
