In An Introduction to MultiAgent Systems, Michael Wooldridge describes multi-agent systems as a paradigm for building and understanding distributed systems where components are autonomous: they control their own behavior while pursuing their own objectives. The emphasis is on many interacting intelligent agents — software processes, robots, or other entities — rather than a single monolithic program. MAS research spans how to design capable individual agents and how to design societies of agents that can work together (or compete) effectively. This course uses that classical framing and later connects it to today’s LLM-based multi-agent applications.

An agent is a system situated in an environment: it receives information through sensors (API responses, user messages, sensor readings) and acts through actuators (tool calls, motor commands, database writes). Autonomy means the agent decides what to do next based on its internal state and goals, not that every step is hard-coded by a human operator. Practical definitions also stress reactivity (responding to changes), pro-activeness (taking initiative), and social ability (interacting with other agents or humans). In LLM stacks, a single chat model with a tool loop is often described as an agent; multiple such processes messaging each other form a multi-agent system.

A single agent (or one orchestrated pipeline) keeps all reasoning and state in one place. That is simpler to debug and deploy when the task is narrow. Multi-agent designs split responsibilities across entities with distinct roles, memories, or policies. Reasons to use MAS include: modularity (teams can own different agents), parallelism (independent subtasks), robustness (failure isolation), specialization (different models or prompts per role), and explicit interaction (negotiation, critique, voting). Costs include coordination overhead, harder testing, and risk of incoherent or conflicting behavior if protocols are weak.

Wooldridge contrasts cooperative settings (e.g., teams, supply-chain partners optimizing a shared objective) with self-interested settings closer to markets, where agents optimize private utility and interaction is structured by prices, contracts, or rules. Real systems are often mixed: engineering squads cooperate inside a firm while competing for budget; LLM agents may cooperate on a user task yet still reflect prompts that encode conflicting sub-goals. MAS theory supplies vocabulary and mechanisms — auctions, voting, norms — for both extremes. Later chapters connect competition to game theory and cooperation to joint plans and commitments.

Organizations adopt MAS-style software when problems are naturally distributed: different owners, data silos, or geographic sites. Scientifically, MAS models social and economic processes (traffic, ecosystems, electronic markets) where no central controller exists. In AI engineering, multi-agent setups can mirror human team workflows — researcher, coder, reviewer — each implemented as a policy with its own context window. Benefits include reuse of specialized components and clearer interfaces between them. Drawbacks: you must engineer message schemas, termination conditions, and failure recovery across process boundaries.

Historically, MAS appeared in robot soccer, sensor networks, manufacturing control, logistics and ride-sharing, simulation of populations, and grid computing. In enterprise IT, workflow engines and microservices often resemble societies of agents even when not labeled as such. Recent LLM applications reuse the same patterns for research assistants, coding crews, customer support triage, and games / simulations. The domain shapes whether you need hard real-time guarantees, formal protocols, or flexible natural-language chatter between agents.

LLM-based agents still perceive (tokens, tool outputs), update internal state (context, memory stores), and act (send messages, call APIs). Classical MAS contributed speech-act taxonomies, negotiation protocols, and commitment models; LLM systems often approximate these with natural language and ad-hoc prompts. Differences: non-determinism and hallucination complicate guarantees; context limits force summarization; cost and latency dominate at scale. Frameworks such as Microsoft’s AutoGen (multi-agent conversation, tools, humans in the loop) operationalize patterns that MAS researchers described abstractly. This course bridges terminology so you can read both academic MAS and vendor docs.

We move from definitions and architectures (Chapters 1–2) to communication and coordination (3–4), then planning and negotiation (5). The second half covers emergence and incentives (6), LLM-native frameworks (7), evaluation (8), safety (9), and production patterns (10). Related material appears in Agentic AI, MCP, and Reasoning & CoT — use those for deeper single-agent reasoning and tooling.