Multi-Agent Interaction Patterns

Method: We grouped 40 cited patterns from foundational and 2026 literature into nine families. Each record names the coordination mechanism, multi-agent rationale, use cases, maturity, and a defining source.

Background

Multi-agent systems add another design layer: coordination. A result depends not only on the models, but on who proposes, verifies, routes work, shares state, and decides. We call that arrangement an interaction pattern.

This survey groups 45 patterns into established, emerging, and frontier work. Each record states the mechanism, why it requires multiple agents, the failure modes it addresses, and a defining source. Maturity is an editorial assessment of evidence and adoption, not a benchmark ranking.

Key findings

1. 1Coordination is a separate design choice. Model selection does not determine who proposes, verifies, routes work, shares state, or makes the final decision. The interaction pattern does.
2. 29 patterns are established. Role-based workflows, independent verification, judging, ranking, debate, and ensembling have defining sources and broad use.
3. 314 are emerging and 22 remain frontier. These categories mark the strength and age of the evidence. They are editorial maturity labels, not performance scores.
4. 4Adaptive topology is a clear recent direction. Several 2026 papers route or rebuild the communication graph by task, query, or round instead of fixing it at design time.
5. 5More agents are not a default upgrade. The MAST study reports that multi-agent systems often fail to outperform a single agent. Added roles need a specific purpose and an explicit coordination cost.

The pattern catalogue

45 patterns across 9 families, grouped by coordination mechanism. Each record states the core dynamic, its multi-agent rationale, the MAST failure categories it may address, example domains, maturity, and a defining source.

Browse all patterns, fields, sources, and CSV/JSON downloads →

The established core

The dataset labels 9 patterns as established. Build-Verify-Reflect separates construction from checking and feeds failures into a bounded retry. Standard operating procedures and role specialisation divide work into named responsibilities.

Other established patterns make evaluation or aggregation explicit. LLM-as-judge and arena rating assign comparison to a separate role. Structured debate tests opposing arguments before a decision, while layered ensembling combines independent proposals.

The emerging tier

The 14 emerging patterns explore coordination that is less standardised in current systems. Adversarial stress testing, multi-model consensus, and negotiation seek genuinely different perspectives or objectives. Evolutionary selection, teacher-student transfer, and competitive arenas use populations, transfer, or incentives to change the result over repeated rounds.

Recent research

The 22 frontier entries focus on graphs that adapt to the query, the current round, or changes in shared state. These are recent research results, often supported by a single paper, so the catalogue treats them as directions to watch rather than settled practice.

CoAgent illustrates the shared-state problem. It assigns a serial order, allows writes to proceed, then notifies affected agents so they can repair conflicts instead of holding a lock across long inference steps.

Where multiple agents help

The MAST study found that multi-agent systems often fail to outperform a single agent, with many failures caused by system design rather than model limits. A second agent is justified when it has a distinct job: independent verification, isolated context, adversarial review, specialised knowledge, or an explicit decision role. Without that separation, extra agents add cost and coordination risk without adding evidence.