In a typical perfect-information strategy game like Chess, an agent can calculate its best move by searching the state tree for the best possible move, while assuming that the opponent will also make the best possible move (i.e. Mini-max).

I would like to use this approach in a "game" modeling economic activity, where the possible "moves" would be to buy or sell for a given price, and the goal, rather than a specific class of states (e.g. Checkmate), would be to maximize some function F of the agent's state (e.g. F(money, widget) = 10*money + widget).

How to handle buy/sell actions that require coordination between both parties, at the very least agreement upon a price?

The cheap way out would be to set the price beforehand, maybe based upon the current supply -- but the idea of this simulation is to examine how prices emerge when freely determined by "perfectly rational" agents.

A great example of what I do not want is the trading algorithm in SugarScape -- paraphrasing from Growing Artificial Societies p101-102:

when a pair of agents interact to trade, they each compute their internal valuations of the goods, then a bargaining process is conducted and a price is agreed to. If this price makes both agents better off, they complete the transaction

The protocol itself is beautiful, but what it cannot capture (as far as I can tell) is the ability for an agent to pay more than it might otherwise for a good, because it knows that it can sell it for even more at a later date -- what appears to be called "strategic thinking" in this pape at Google Books Multi-Agent-Based Simulation III: 4th International Workshop, MABS 2003... to get realistic behavior like that, it seems one would either (1) have to build an outrageously-complex internal valuation system which could at best only cover situations that were planned for at compile-time, or otherwise (2) have some mechanism to search the state tree... which would require some way of planning future trades.

Note: The chess analogy only works as far as the state-space search goes; the simulation isn't intended to be "zero sum", so a literal mini-max search wouldn't be appropriate -- and ideally, it should work with more than two agents.