With multiple agents, individual rationality does not guarantee global feasibility: double booking, conflicting tool calls, duplicated work, or deadlock waiting on replies. Coordination selects who does what, when, and with which resources. Cooperation assumes aligned interests; competition needs rules and incentives. Your stack must expose allocations (who owns a task), synchronization points (barriers), and conflict resolution (priority, auction, human arbiter).

The contract net protocol: a manager announces a task; workers submit bids (capability, cost, time); the manager awards a contract. Variants power logistics and cloud schedulers. For LLM agents, bids might encode confidence, estimated tokens, or required tools. Keep awards logged so you can explain why agent B won. Watch for collusion or herding if bids are just paraphrases of the same LLM policy.

A central allocator sees more context and can enforce policy but risks bottlenecks. Decentralized agents pick work from a queue using local rules (skill tags, load). Many production systems use a hybrid: orchestrator for critical paths, workers pulling from a prioritized backlog otherwise. Encode task state (OPEN/CLAIMED/DONE) in a store all agents respect.

For decisions that are not decomposable (pick architecture A vs B), agents may vote or score options. Simple majority can ignore minorities; weighted votes reflect expertise or risk ownership. In LLM ensembles, aggregation might be majority tool choice or a meta-critic ranking plans. Document quorum rules and tie-breakers (human, risk-averse default, roll back).

Research on teamwork (e.g., Cohen & Levesque on joint intentions, SharedPlans formalisms) stresses mutual belief, commitment to the joint goal, and helping when teammates fail. Practically, represent a team task graph with explicit dependencies and a fallback owner for each node. LLM role-play is not enough — persist the graph where all agents read the same truth.

Beyond protocols, agents follow norms: obligations, prohibitions, sanctions. Implement norms as policy-as-code checked before tool execution (OPA-style) or as prompt constitutions backed by validators. Multi-agent systems amplify policy errors — one rogue agent can trigger many downstream calls. Centralize authorization for sensitive acts even if planning is distributed.

Sometimes coordination is indirect: agents leave traces in an environment others react to (stigmergy). Useful for exploration and creative search; risky for compliance tasks. Engineered coordination uses explicit workflows. Blend them: allow emergent ideation inside a bounded sandbox, then funnel winners through a gate with schema validation and human approval.

You now have architectures, communication, and coordination patterns. Chapter 5 dives into multi-agent planning and negotiation when tasks interact, resources clash, and information is private. Bring your protocol diagrams and allocation rules — they constrain which plans are even legal.