Separate the agentic harness from the LLM subscription to remove the conflict of interest

This harness is built and shipped by the same company that sells the cloud LLM subscription it consumes. That creates a conflict of interest. The company that sets the price of tokens and the monthly quota is also the company that builds the tool that spends them. Because the same party earns more when more tokens are consumed, a user cannot be sure that the design of the harness is made for the user's benefit rather than for the vendor's revenue.

This is not a claim that any specific number is fake. It is a structural problem. When one company controls both the meter and the machine that runs on it, the user has no independent guarantee that token use is kept honest. We can see the difference by comparing with harnesses that do not sell a subscription. Tools like the PI coding agent, which only build the software and never sell the model, tend to use fewer tokens for the same task. That suggests the incentive matters. A vendor that only sells the tool is pushed to make it efficient, while a vendor that also sells the tokens is pushed the other way.

I am asking for a clear separation between the two roles. The agentic and harness software should be developed and shipped apart from the sale of the LLM subscription, so that the people who build the tool are not the same people who profit from how many tokens the tool burns. With a clean barrier between them, the conflict of interest disappears and users no longer have to suspect that heavy token consumption is planned. If a full separation is not possible, then at minimum the harness should show the user a clear and honest token cost for every run and for every parallel sub-agent, so the consumer always knows what they are paying.