Distributed Artificial Intelligence

Distributed artificial intelligence (DAI) is the subfield of AI that studies how several agents or nodes, spread out and without a single central controller, cooperate —or compete— to solve problems that exceed any one agent. Rather than concentrating all reasoning in a single system, DAI distributes knowledge, decisions and computing power across entities that communicate, coordinate and, at times, negotiate with one another.

Each of those autonomous agents is covered in the intelligent agents entry; the focus here is the level above: how they are organized as a whole. The discipline took shape in the 1980s and 1990s, and collections such as Readings in Distributed Artificial Intelligence (1988), edited by Alan Bond and Les Gasser, fixed its basic vocabulary.

Two roots: distributed problem solving and multi-agent systems

Bond and Gasser drew a distinction that still holds. Distributed problem solving starts from a global task and splits it among nodes that work in parallel toward a shared goal; the challenge is decomposing the problem well and reassembling the partial solutions. Multi-agent systems start from autonomous agents with goals of their own that must interact: here communication, cooperation and coordination —and even conflict— take center stage. To allocate tasks without a central boss, Reid G. Smith introduced the Contract Net Protocol in 1980, a scheme of announcements and bids that still informs today's auction mechanisms and negotiation protocols. Game theory supplies the framework for reasoning about incentives when agents do not share goals, as Gerhard Weiss compiled in his 1999 reference volume.

Not the same as distributed computing

DAI should not be confused with plain distributed computing. Splitting a computation across many machines —or training a model on several GPUs— distributes the work, but the nodes decide nothing: they run what they are assigned. In distributed artificial intelligence, by contrast, each agent keeps some autonomy to decide, prioritize or negotiate. The difference is not the hardware but who makes the decisions, and how independently.

The 2026 revival: LLM multi-agent systems and federated learning

After years in the background, DAI is back at the center. Multi-agent systems built on large language models assign roles to several models —to plan, critique, execute— that collaborate on a single problem, reviving the old questions about coordination. In parallel, federated learning, formulated by Brendan McMahan and colleagues at Google in 2017, trains a shared model without centralizing data: each device learns locally and shares only updates. None of this is settled. Recent surveys show that adding agents does not guarantee better results than a single well-designed agent, and that reliable coordination and the emergence of collective behavior —useful or unwanted— remain open problems.