Yes, every agent based model that eats is like Darwinbots
Well, are you wanting to model real chemistry or fake chemistry? If you're interested in the actual physical chemicals used by a life form (wether carbon or no) you're going to need a super computer.
I can make a simulation which pretends that some of its organisms are silicon life forms (which BTW isn't all that likely chemically speaking. You're more likely to see critters that use carbon but use otehr polar molecules instead of water, probably in either very cold or very dense environments.) but that's not going to tell me anything about silicon life forms.
Point is you're going to need a simplified biochemistry whatever your plans are. And BTW DB's biochemistry, derived from bit patterns etc. isn't necessarily an abstraction of actual biochemistry. There is no analogue to citric acid. There are no explicit nitrogen cycles. It's purely an absraction of what biochemistry inherantly does, which is rearrange bits in macromolecules.
Now, as to your idea:
It's theoretically sound, assuming a very good AI model. However you might find it to be computationally intractible. Specifically, you'll probably be needing to do some rather intensive rigid body dynamics and collision detection, on the order of several thousand traingle primitives at least. Due to the dynamic nature of the agents, you're not going to be able to use BSPs to simplify collision detection between agents.
And as to the "chemistry" of it, you're not really providing an energy flow, which is key to understanding the dynamics of ecosystems. You're primarily exchanging parts between agents, which would be a kind of Mad Max system dynamic, with all the agents scavenging off each other.
It might be a good idea to provide some inherant way to simulate energy flow. Perhaps being able to construct triangles with energy and destruct triangles into energy.