Personally the term agent is too overloaded. An agent is the loop. That is it, this loop handles context management and continuity of the calls to the LLM. The LLM is a STATELESS REST API call. The harness is the tooling and ways to interact in with the local system for the model. When you start looking at the boundaries this presents, you get a more accurate reprensentation of what is happening. It also decouples the agent loop from the specific model. Model's become more interchangeable reasoning service calls.
Good points, and I agree. There is a reflex out there right now that anything generated by LLM's is not human attributable. However, the human is the initiator, the guide, the decision maker and judge. The LLM is an amplifier of the human's thought. If we keep the full working session history we can show the human contribution explicitly, but we also expose the full thought processes, etc, and it becomes a privacy question if we were to attach that information to the output product. How do we shift the perceptions that are driving these hostile viewpoints?
I propose Channel Engineering, and will be releasing a paper with the full reasoning soon. The TLDR is that it is a communication channel between a Human and AI, mediated by a deterministic agent loop process. That loop is in control of the channel and needs to own it completely to solve for the reliability of the delivered outcomes in ways that can be concretely measured.