← All chapters
Multi-Agent Systems & Orchestration
Coordinating multiple agents — roles, communication, and orchestration patterns.
20 terms
Agent CommunicationAgent communication refers to the protocols and data formats through which agents in a multi-agent system exchange information, task descriptions, intermediate results, and control signals. Communication can be synchronous — one agent blocks waiting for another's reply — or asynchronous, with agents posting messages to a shared queue. Well-defined communication contracts prevent agents from misinterpreting each other's outputs and simplify debugging when pipelines fail.Agent DebateAgent debate is a pattern in which two or more agents argue opposing positions on a question, critique each other's reasoning, and converge on a shared conclusion through structured dialogue. The adversarial dynamic surfaces assumptions, identifies weaknesses in arguments, and tends to produce more thoroughly reasoned outputs than a single agent working alone. Debate is especially useful for complex decisions where a single model's bias might otherwise go unchallenged.Agent HandoffAn agent handoff is the controlled transfer of task ownership from one agent to another within a multi-agent pipeline. The transferring agent packages current progress, outstanding objectives, and relevant context into a structured payload that the receiving agent can immediately act on. Handoffs require well-specified interface contracts: if the receiving agent misinterprets the payload, the task can stall or silently degrade without an obvious error signal.Agent NetworkAn agent network is a collection of agents connected by communication links that allow any agent to invoke or message any other, without necessarily a central orchestrator. Network topologies range from peer-to-peer graphs to hub-and-spoke arrangements. Agent networks enable emergent collaboration but make control flow harder to reason about; explicit routing rules, message schemas, and observability tooling are essential to prevent chaotic or looping interactions.Agent SwarmAn agent swarm is a large collection of lightweight, often identical agents that work in parallel on decomposed subtasks without centralized coordination. Behavior emerges from the aggregate of individual contributions rather than from an explicit orchestration plan. Swarms excel at embarrassingly parallel workloads — such as batch document processing or exhaustive search — but require careful result aggregation and are harder to debug than tightly orchestrated pipelines.AggregationAggregation is the step in which outputs from multiple parallel worker agents are combined into a single coherent result. Depending on the task, aggregation may be as simple as concatenating text sections, or as complex as running a separate synthesis agent that resolves conflicts, removes duplication, and ensures consistency across contributions. Poor aggregation can introduce contradictions or lose important contributions from individual workers.Blackboard ArchitectureIn a blackboard architecture, all agents read from and write to a shared data structure — the blackboard — rather than communicating directly with each other. Any agent can post a partial result, and any other agent whose expertise applies to that result can pick it up and contribute further. Originally developed for AI planning systems, the pattern works well for tasks where different specialists must contribute opportunistically to a common evolving solution.ConsensusConsensus is the process by which multiple agents converge on a shared answer, plan, or decision after independently reasoning about a problem. Mechanisms include majority vote among agent outputs, iterative debate until agents agree, and weighted averaging of confidence scores. Consensus mechanisms can significantly reduce the variance and error rate of final outputs at the cost of additional model calls and latency.Cooperative vs Competitive AgentsCooperative agents share a common objective and coordinate to achieve it, dividing labor and sharing information freely. Competitive agents have misaligned or opposing objectives — as in debate, red-teaming, or game-playing — and each tries to maximize its own outcome. Production multi-agent systems are usually cooperative, but deliberately competitive sub-systems such as adversarial reviewers or red-team agents can improve overall output quality by stress-testing cooperative agents' work.DelegationDelegation is the act of an agent passing responsibility for a task — or a portion of a task — to another agent. The delegating agent specifies the objective, any constraints, and the expected output format, then waits for or polls the result. Effective delegation requires trust that the receiving agent will complete the work correctly; in practice, the delegating agent often validates returned results before incorporating them into its own output.Hierarchical AgentsHierarchical agents are organized in a tree-like structure where higher-level agents decompose goals and delegate to lower-level agents, which may in turn spawn further sub-agents. This mirrors organizational hierarchies: a top-level planner handles strategy, middle-tier agents manage subsystems, and leaf agents execute atomic actions. The structure scales well to complex tasks but introduces depth-related latency and makes end-to-end failure attribution more difficult.Manager-Worker PatternThe manager-worker pattern divides agents into two tiers: a manager that holds the task plan and assigns work items, and workers that execute those items and return results. The manager tracks which items are complete, handles retries on failure, and merges results into a final output. This pattern is a simplified form of hierarchical agents, typically without deep nesting, and maps naturally to batch processing pipelines with homogeneous subtasks.Message PassingMessage passing is the mechanism by which agents communicate by sending discrete, structured messages to one another rather than reading a shared mutable state. Each message typically carries a sender identity, a recipient, a message type, and a payload. Message-passing architectures are easier to trace and debug than shared-state systems because every inter-agent interaction is an explicit, logged event rather than a silent mutation of global memory.Orchestrator AgentAn orchestrator agent is a coordinating model responsible for decomposing a high-level goal into subtasks, assigning those subtasks to specialist worker agents, and synthesizing their outputs into a coherent final result. The orchestrator does not itself execute low-level actions; instead it plans, delegates, monitors progress, and handles failures by re-routing or retrying. This separation keeps individual agents focused while concentrating planning complexity in one place.Role AssignmentRole assignment is the process of defining distinct identities and responsibilities for each agent in a multi-agent system. Each role specifies what kinds of tasks the agent handles, what tools it may use, and how it should communicate with other agents. Clear role assignment prevents agents from duplicating effort, establishes accountability for each pipeline stage, and makes system behavior more predictable and auditable.Routing AgentA routing agent classifies incoming requests and dispatches them to the most appropriate specialist agent or pipeline without itself attempting to answer the request. Routing decisions can be rule-based, model-driven, or a hybrid. Effective routing reduces latency and cost by preventing generic agents from handling queries better served by specialists, and it isolates failures — a routing error affects only the misdirected request, not the whole system.Shared Memory (Multi-Agent)Shared memory in a multi-agent context is a common store — such as a database, file, or in-process object — that all agents in a system can read from and write to. It enables agents to coordinate without direct message passing but introduces concurrency risks: two agents writing simultaneously can corrupt state, and one agent reading stale data can make poor decisions. Careful locking, versioning, or conflict-resolution strategies are necessary in production systems.SpecializationSpecialization is the design principle of configuring each agent in a multi-agent system to excel at a narrow category of tasks — such as code generation, web research, or structured data extraction — rather than attempting everything with a single general agent. Specialized agents can be tuned with targeted system prompts, given only the tools they need, and evaluated on a focused benchmark. Specialization improves quality and cost efficiency but requires more upfront system design.Supervisor PatternThe supervisor pattern is a multi-agent architecture in which a supervisor model routes each incoming task to the most appropriate worker and reviews worker outputs before returning a final response. Unlike a simple orchestrator that sequences workers linearly, a supervisor can dynamically select workers, ask workers to revise their output, or escalate difficult tasks. The pattern balances flexibility with control and is a common choice for production customer-facing systems.Worker AgentA worker agent is a specialized sub-agent that receives a focused subtask from an orchestrator, executes it using its own tools and reasoning, and returns a structured result. Workers are typically narrower in scope than orchestrators — a worker might only handle web search, code execution, or data extraction — which allows deeper specialization and simpler evaluation. The same worker type is often instantiated multiple times in parallel to process independent subtasks concurrently.