Re: AGENTS.md

On Thu, Aug 6, 2026 at 1:43 PM Eli Zaretskii <[email protected]> wrote:

> Once again, it is well known that not all LLMs automatically read a
> file by this name.

Can we please apply the policy of using accurate terminology to the
area of LLM-assisted development?

* LLM (Large Language Model): A collection of statistical data derived
from a training corpus.
* Provider: A software package that accepts prompts, inferences the
Model and returns responses, in a machine-readable format through a
well-defined API.
* Harness: A software package that combines prompts from the user’s
request, the user’s configuration, the project, and the harness
itself, sends them to the Provider, and exposes a set of tools so that
the Agent can read and modify files and perform actions on the user’s
behalf.
* Agent: The system as a whole, comprised of the Model, the Provider,
and the Harness.

Sean said he tested a commercial LLM; but the reading of AGENTS.md and
relaying it to the LLM is the responsibility of the harness.

(I get the impression that Jean had started to use LLMs before the
concept of a harness crystallized, and might be somewhat skewed in his
assessment of how people at large use LLMs for coding.)