Preliminary requirements

• It should be played over a number of turns, and without some definite ending point.
• Discrete interactions should involve a low number of actors (if possible)
• Interactions should be described as normal form games, but with a payoff matrix that just lists game effects
• There should not be a pre-defined utility function - agents should determine their own utility functions
• The environment should be dangerous enough that agents which make poor decisions are eventually eliminated
• Even though there are no pre-defined utility functions, there should be some loose measure of 'how well you are doing' which should reflect how far away you are from being eliminated, and how much you are able to alter the environment to suit yourself.
• There should be some concept of location, or territory, which will constrain agent interactions and communication
• There should be a standard language used to describe the environment, and perform social interactions.
• Agents should be able to identify other agents uniquely (this is not an absolute requirement, but I think it is a good one to start with)
• There should be some objects/tokens which can be exchanged between agents. Perhaps start with a limited set, but some concept of money is extremely useful in doing negotiations.
• There's more, but that's enough for now.

Thoughts towards fulfilling these requirements

I like the idea of using food to determine when an agent is eliminated. That is to say, there are food tokens, and each agent must consume a number of food tokens every time period, or have their performance degrade, and eventually be eliminated.

As far as location goes, I think using a location/path (vertex/edge) model like MUDs is a good idea, although it does not allow for pursuit very well.

I prefer a notion of continuous time to discrete time. In other words, the game does not proceed according to discrete turns in which everyone gets an oppourtunity to act. Rather, each action has a duration in terms of a number of turns it takes to perform. We can give communicating a cost to perform, and so avoid the infinite meeting problem of agents trying to achieve a consensus on what actions to perform as a group. I like this, as it attaches a natural cost to communicating which still encourages communication (assuming we make the cost of communicating low relative to actions which change the game state).