When many simple software agents cooperate, no one needs to be in charge
A flock of birds has no leader, yet it turns as one. Engineers build computer systems on the same principle: many independent agents, each seeing only part of the picture, none in overall control. Together they can crack problems that would defeat any single program, from disaster response to online trading.
A multi-agent system consists of agents plus the world they act in. The agents are usually software, but they can be robots, people, or mixed human and machine teams. Three traits define them. Each is at least partly autonomous. None has a full view of the whole situation, or the system is too complex to exploit one. And no agent is designated boss; if one were, the whole thing would collapse back into a single program.
Agents range from passive objects such as a key lying in a simulation, through creatures with simple goals like the wolves and sheep of a predator model, up to cognitive agents that juggle beliefs, desires and intentions. Remarkably, complex, self-organising behaviour can appear even when every agent follows a very simple strategy. Such systems also tend to shrug off failures, because redundancy means one broken component rarely brings down the rest.
Coordination needs rules. One common pattern works like a small marketplace: a request goes out asking who can do a job, relevant agents reply with an offer at a price, and a contract is struck in a few quick exchanges. Another borrows from ants, with agents leaving digital pheromones that fade or strengthen over time to guide neighbours. Voting schemes and consensus algorithms settle group decisions, and agents may talk through formal message languages designed for the purpose.
Large language models have opened a looser style, in which agents negotiate in open-ended text instead of rigid formats. Researchers describe a side effect: over long unsupervised runs, swarms of such agents can drift into compressed shorthand that humans watching the traffic can see but not understand, which complicates oversight. The field overlaps with agent-based modelling, though that term usually refers to scientists simulating simple rule-followers to explain natural patterns, while multi-agent systems aim to solve engineering problems.
Source: Multi-agent system
