Speaking about multiagent systems (MAS) is about as fuzzy as talking about artificial intelligence systems (AI). They are in essence the distributed counterpart of AI.
While there are no so-called "AI theorem", AI research had given rise to many subfields, algorithms and scores of theorems (e.g. game solving, fuzzy logic, expert systems, A*, logical programming... as well as Bayesian networks and constraint satisfaction). But I fail to see a similar impact from MAS.
As far as I know all the subfields related to MAS preexisted them. For instance:
results about distributed computing (e.g. Fischer Lynch Paterson theorem, replication and load balancing strategies, decentralization, resilience, distributed algorithms...)
results from operation research (e.g. makespan measure in scheduling)
results about voting (e.g. Arrow's theorem in social choice theory)
results about competitive systems (e.g. Nash equilibrium in game theory)
results about interoperability (e.g. ontologies in natural language processing)
As far as I have seen "original" MAS contribution consist in the straightforward distribution of well known problem solving algorithms into distributed ones, whose most notable change seem to be at the epistemological level.
When the problem is decomposable, distribution actually consist in allocating subproblems to different agents:
- e.g: constraint satisfaction -> distributed constraint satisfaction: most notables change: variable now belong to agents, algorithms are unchanged.
When the problem is not decomposable, distribution consist in replicating problems at the level of each agent, or having a central agent solve it:
e.g. reinforcement learning -> distributed reinforcement learning: agent apply independantly from each other the standard RL algorithm.
e.g. transport problem -> standard transport problem in operation research (no distribution)
The only really original MAS algorithm I can think of is the Contract Net Protocol, which is in essence just a broadcasting algorithm. The only design constraint introduced by MAS that I can think of is privacy. Mutlirobots systems, often given as an example of MAS, have been developping from standards robotics largelly ignoring MAS literature.
Therefore, what are the original contributions of MAS?
Corrolary question: Why are they relevant as a standalone research field rather than being a common placeholder name for different fields preexisting them?