"I optimized my own reward beautifully. The other six agents did exactly the same thing, and now the warehouse is on fire."
An Agent That Maximized Locally and Regretted Globally
The first five parts distributed the machinery of one model; Part VI distributes the decision-making itself, replacing the single learner with a population of autonomous agents that each perceive, choose, and act, and that must coordinate to behave as one system. Every earlier part split a computation that a single coordinator could, in principle, have run alone: the gradient, the shard, the request. The intelligence axis is different in kind. Here there is no global loss to average and no central scheduler with the full picture; each agent holds its own objective, its own partial view, and its own thread of control, and the system-level behavior is whatever emerges when those local choices collide. This part gives you the foundations, the game theory, the architectures, the learning algorithms, and the orchestration patterns that turn many self-interested or merely independent actors into a coherent distributed intelligence, and it shows where that coherence is fragile.
Part Overview
Distribution has meant one thing for five parts: take a workload that overflowed a single machine and spread it across many, paying communication and failure taxes to buy capacity. The data, the training, the model, the inference, all of it was ultimately one computation a single powerful enough node could have performed, sliced for throughput and memory. Part VI distributes the last and strangest axis, intelligence itself, and the slice is not merely physical. When decision-making is distributed, there is no single objective being minimized and no coordinator holding the complete state; there are many agents, each with its own goals, its own observations, and its own actions, and the question shifts from "how do we split this computation correctly" to "what happens when many autonomous deciders interact, and how do we steer the result."
That shift reframes every tool the book has built. Communication is no longer a tax on an otherwise-exact computation; it is the medium through which agents negotiate, signal, and form agreements, and its absence produces not a slower system but a different equilibrium. Failure is no longer a node that stops; it is an agent that defects, or a peer whose incentives diverge from yours. The performance metrics of Chapter 5 gain companions drawn from economics and control: social welfare, regret against an adversary, convergence to an equilibrium, the price of anarchy. Part VI assembles these ideas in order, moving from the conceptual foundations of distributed artificial intelligence, through the game-theoretic language that makes "self-interested agents" precise, into the engineering of multi-agent systems, the learning dynamics of multi-agent reinforcement learning, the emergent coordination of swarms, and finally the orchestration of large-language-model agents into pipelines that solve tasks no single agent could.
The narrative thread the book has followed returns here transformed. The distributed reinforcement-learning infrastructure of Chapter 20, where many actors fed experience to one learner optimizing a single policy, becomes in Chapter 30 a setting where each actor is also a learner with its own policy and its own reward, so the environment every agent faces is non-stationary because the other agents are changing too. The collective-communication primitives of Chapter 4 reappear as the message-passing substrate of agent coordination, but now the messages carry intent rather than gradients. Read this part as the book's answer to its own final axis: once you can distribute data, training, models, and inference, the remaining frontier is distributing the choosing, and that frontier is where distributed systems meet game theory, control, and the engineering of agents that act in the world.
Every prior part distributed a computation with a single correct answer that a coordinator could, in principle, have verified; Part VI distributes the objective itself, so correctness is replaced by equilibrium, and the central engineering question becomes whether the local choices of many autonomous agents compose into the global behavior you actually want. This is why game theory enters as foundational rather than decorative: it is the only language precise enough to predict what self-interested deciders do, and to design the incentives, protocols, and learning rules that make their collective behavior align with the system's goal. Carry this question into every chapter that follows, what does each agent optimize, what does it observe, and what equilibrium do those local optimizations reach, and you will read multi-agent systems, multi-agent reinforcement learning, swarms, and agent orchestration as five answers to one design problem.
Part Roadmap
- 27 Distributed Artificial Intelligence The conceptual foundations of the intelligence axis: what makes an agent autonomous, why distributing decision-making differs from distributing computation, and the spectrum from cooperative to self-interested that the rest of the part formalizes.
- 28 Game-Theoretic Foundations for Multi-Agent AI The language that makes "self-interested agents" precise: strategies, payoffs, Nash and correlated equilibria, mechanism design, and the price of anarchy that quantifies how far local optimization drifts from the global optimum.
- 29 Multi-Agent Systems The engineering of coordination: agent architectures, communication protocols, negotiation and contract nets, task allocation, and consensus among peers with partial views and no shared scheduler.
- 30 Multi-Agent Reinforcement Learning Learning when every other agent is also learning: the non-stationary environment, centralized-training-decentralized-execution, credit assignment across a team, and the algorithms that reach stable joint policies.
- 31 Swarm Intelligence and Collective Behavior Coordination without coordination: how simple local rules over many minimal agents produce global structure, from ant-colony and particle-swarm optimization to flocking, stigmergy, and robustness through redundancy.
- 32 Distributed Agent Orchestration The modern payoff: composing large-language-model agents into pipelines, with planner-executor patterns, tool use, shared memory, and the routing and fault handling that make a team of agents solve what one cannot.
Read the six chapters in order and the part builds one argument from concept to system. Chapter 27 establishes why distributed decision-making is a distinct discipline; Chapter 28 supplies the formal language; Chapter 29 turns that language into architectures and protocols; Chapter 30 adds learning to those architectures; Chapter 31 shows the opposite extreme, where intelligence emerges from agents too simple to negotiate; and Chapter 32 lands the whole arc in the agent systems practitioners build today. The case studies of Chapter 39 on multi-agent robotics and Chapter 40 on agentic applications draw directly on this part.
What's Next?
Part VI completes the six axes of distribution: with intelligence itself spread across coordinating agents, the book has shown how data, training, models, inference, coordination, and decision-making each scale out. Part VII: Cluster, Edge, and Reliable Infrastructure turns from what we distribute to where it runs and how it stays running. Every system in Parts II through VI assumed machines that could be scheduled, reached, and trusted; Part VII supplies those assumptions, the cluster schedulers that place the work, the edge and fog tiers that push it toward the data, and the reliability and security engineering that keeps a distributed intelligence correct under failure and adversaries. The agents you learned to coordinate here must ultimately be placed on real hardware across real networks, and that placement is where Part VII begins.