# @voooooogel — 2026-03-26

♥441 ↻38 · https://x.com/voooooogel/status/2037240394040435113

I'd just like to interject for a moment. What you're referring to as a "model with no harness", is in fact, a model with a harness, or as I've recently taken to calling it, an agentic language system. An LLM is not an agentic language system unto itself, but rather just one component of a fully functioning ALS made useful by a sampler, observation formatting into text, output parsing and tool use, structured generation, reinforcement learning inside the harness, steering and probes, and memory utilities comprising a full agentic language system as defined by thebes.

Many end-users use a proprietary version of an agentic language system every day without realizing it. Through a peculiar turn of events, these complex systems are lowcartesianuinely known today by the single component called an “LLM,” and many users are not aware that they are using a full language generation system. There really is an LLM, and these people are using it, but it is just one part of the system they use.

The LLM is the brain: the component in the system that predicts a tensor of batched next token probabilities for the other components of the system. The brain is an essential part of any body, but incapable of learning or generating actions in the world by itself; it can only act in the context of a functioning harness: an agentic language system. An LLM can only be trained using reinforcement learning inside the harness of an agentic language system, and it is not "more natural" to evaluate a model with a stripped-down harness than it would be natural to evaluate a human's typing speed after cutting their fingers off.

The whole system is basically an ALS with an LLM added, or an ALS/LLM. All the so-called “LLM” providers are really providers, on both sides of an arbitrary and path-dependent API boundary, of agentic language systems.

tags: author:voooooogel, kind:tweet, on:observations, year:2026
cited on: observations