This needs to be stressed. Reach can be very accurate both RELATIVELY and AUTONOMOUSLY.
For it to be RELATIVELY accurate, it needs to adjusted or based on a known point.
For it to be AUTONOMOUSLY accurate, it does not need to be adjusted or based on a known value.
For both operations, you will get cm results, always relative to your base station. It is just that in the case of RELATIVELY accurate surveys, you will be able to coordinate the same place in space with the same value each time you survey. AUTONOMOUS (other survey institutions will use words like deferentially un-corrected or localized) surveys mean that each time you try and measure a point, you will get a different value for it, varying a few meters.
It should be clear, this is not a downfall of Reach. Even the top dollar systems do this system. Part of survey training is to understand how this works and how to decide if the impact of you survey requires it to be based on a known point or if it can be localized.
Weston, from my understanding of agricultrual systems (which is very limited), they do not need to be based on a true WGS84 or known value. As I understand it, they are localised, so if there is a system in place in the field, you will need to compensate for it by measuring on to it. But I think this is unlikly. If the equipment is just required to drive in a straight line, then using the Reach in an AUTONOMOUS survey mode will be sufficient.
Remember, there is nothing stopping you from making your own localised survey repeatable by placing the base on the same position each time you survey and using the same coordinates each time. There is a reason why the base remembers your last coordinates. Using this methodology will give you repeatable and consistently accurate results.
Weston, if you have to base it on a known point, you can really push the reach to get a decimeter value at 50km from a base station. While it will never give you a fix(if it does, it will be junk data), but you will be pretty much close to the value. The other option is to make a series of 8km baseline surveys and move the base each time to the previously surveyed value. I would not recommend this approach as you can introduce big errors and there is no way of checking your final answer.