Multi-Agent Patterns

Choose the pattern after the split is justified

Pattern choice is a second-order decision. The first-order decision is whether one agent, a fixed workflow, or a traditional program can solve the problem with less ambiguity. Once a split is justified, the pattern describes how responsibilities, context, and control move through the system.

Anthropic's practical guide separates fixed workflows from agents that decide their own process. OpenAI's guide separates single-agent loops, manager-style multi-agent systems, and decentralized handoffs. Research frameworks such as AutoGen, CAMEL, MetaGPT, and multiagent debate explore specific collaboration styles inside that larger design space.

Orchestrator-worker

The orchestrator-worker pattern has a lead agent that interprets the request, decides which subtasks exist, delegates work to specialized workers, and synthesizes the returned artifacts. It is the natural pattern for work where the subtask list cannot be known before the input arrives.

Research is the cleanest fit. A lead agent can inspect the question, create separate investigations, and ask workers to search different source classes. The lead then merges the findings and checks coverage. Anthropic's Claude Research system is an explicit production example of this shape.

The danger is that orchestration can become theater. If the lead agent cannot evaluate worker output, or if workers return prose instead of verifiable artifacts, the system has merely distributed uncertainty across more places. The coordinator must own acceptance criteria, not just task assignment.

Pipeline

A pipeline is a fixed sequence: classify, retrieve, draft, verify, format; or plan, implement, test, summarize. It may use multiple model calls or multiple named agents, but the order is predetermined. This is often easier to test than a dynamic multi-agent system because each stage has a stable contract.

Pipelines are appropriate when quality improves because each step narrows the problem. A writer agent should not also be the policy classifier if policy mistakes are expensive. A code generator should not also be the only reviewer if known regressions require a checklist. Fixed stages create audit points.

A pipeline is not automatically less sophisticated than an autonomous swarm. It is often more mature because the system designer admits which choices should remain deterministic. Use pipelines when you can define the steps, the artifacts, and the gates before runtime.

Debate and ensemble

Debate and ensemble patterns run multiple attempts or perspectives and combine them. The simplest ensemble asks several agents to answer independently and then votes, ranks, or judges. Debate adds rounds where agents critique one another before a final answer is selected.

Du et al.'s multiagent debate paper is a canonical reference here: multiple model instances propose answers and reasoning, then debate over rounds to improve reasoning and factuality in the studied tasks. This pattern is most plausible when errors are diverse and independent enough that disagreement is signal, not noise.

The trap is false confidence. Three agents can agree because they share the same model blind spot, the same retrieved source, or the same flawed prompt. A debate system still needs external evidence, a judge rubric, and a rule for when disagreement should trigger more research rather than majority vote.

Swarm and hierarchy

"Swarm" usually means a larger group of agents coordinating through local rules, shared messages, or emergent interaction rather than a single fixed sequence. In LLM systems, the term is often used loosely, so treat it with caution. Ask what state is shared, who can call tools, and how work stops.

A hierarchy is more concrete. Agents are arranged in levels: a top-level manager delegates to domain managers, who delegate to workers. This can reduce local complexity, but it also increases the risk of lossy summaries and unclear accountability. Every layer should compress artifacts, not hide decisions.

CAMEL is important historically because it studies role-playing communicative agents as a way to investigate cooperation. MetaGPT is important because it applies structured human workflows and role assignment to reduce the cascading hallucination risk of naive multi-agent chaining. Together, they show both the ambition and the fragility of more social agent designs.