Multi-Agent Systems

A multi-agent system (MAS) in artificial intelligence is an architecture comprising two or more autonomous agents — each capable of perceiving its environment, reasoning about a goal, and taking actions — that communicate and coordinate with one another to accomplish tasks that are too complex, broad, or parallelisable for a single agent to handle effectively. In contemporary AI, the term most commonly refers to systems in which large language models (LLMs) serve as the reasoning engines powering individual agents, with each agent assigned a distinct role or specialisation, and an orchestrating agent or workflow structure coordinating their interactions.

Historical Context

The study of multi-agent systems has roots in the distributed artificial intelligence literature of the 1970s and 1980s, where researchers investigated how communities of problem-solving programs could cooperate and negotiate to accomplish objectives beyond any individual program's competence. Classical MAS frameworks drew on game theory, distributed computing, and organisation theory to model agent incentives, communication protocols, and coordination mechanisms.

The contemporary resurgence of multi-agent systems in AI has been driven by the development of capable LLM-based agents from 2022 onwards. Demonstrations such as AutoGPT and BabyAGI in early 2023 showed that LLMs could iteratively plan and execute multi-step tasks using tool access, sparking significant research and commercial interest. Frameworks such as AutoGen, CrewAI, and LangGraph subsequently provided structured primitives for defining agent roles, communication patterns, and execution graphs, making multi-agent architectures more tractable for production deployment.

Core Components

Agents

Each agent in a multi-agent system wraps an LLM with a specific system prompt that defines its role, a set of tools it can invoke (such as web search, code execution, database queries, or API calls), and optionally a memory store for retaining information across turns. The agent operates a reasoning loop — typically based on the ReAct pattern (Reason + Act) or similar frameworks — in which it alternates between producing thought traces and taking tool actions until it completes its assigned subtask or signals that it requires input from another agent.

Orchestration

An orchestrator — itself often an LLM-powered agent — receives a high-level task, decomposes it into subtasks, assigns subtasks to appropriate specialist agents, monitors their progress, and integrates their outputs into a final result. Orchestration patterns include hierarchical delegation (a supervisor assigns tasks to workers), sequential pipelines (agents pass outputs forward in a chain), parallel fan-out (multiple agents work simultaneously on independent subtasks), and debate or consensus architectures (multiple agents independently produce answers that are then synthesised or voted upon).

Communication and Memory

Agents in a multi-agent system communicate through structured messages, shared state objects, or memory stores. Short-term memory — the context window of each agent — holds the immediate conversation and recent tool results. Long-term memory may be externalised into a vector database or key-value store, enabling agents to retrieve information from prior sessions or to share findings with other agents without passing everything through the context window.

Applications and Demonstrated Impact

Multi-agent systems have demonstrated significant performance improvements over single-agent approaches on complex benchmarks. In software engineering, agentic systems comprising a planner, a coder, and a code reviewer agent have achieved better performance on the SWE-bench repository-level coding benchmark than any single agent in isolation. In cybersecurity, incident response pipelines using multi-agent coordination have reduced mean time-to-investigate from 30 minutes to 30 seconds in reported deployments.

Enterprise adoption has grown substantially: a 2025 survey by MIT Sloan Management Review and Boston Consulting Group found that 35% of organisations surveyed had adopted AI agents, with a further 44% planning to do so. The multi-agent systems market is projected to reach US$184.8 billion by 2034, according to Dimensions Market Research.

Applications span software development (code generation, review, testing, and deployment agents), research and analysis (literature review, data collection, and synthesis agents), customer service (routing, resolution, and escalation agents), supply chain management, and scientific discovery.

Challenges and Limitations

Multi-agent systems introduce engineering complexity that single-agent architectures avoid. Debugging is harder because errors can originate at any agent boundary and propagate through the system in non-obvious ways. Agent communication failures or hallucinations can cascade, producing outputs that are confidently wrong and difficult to trace. Context management across many agents raises token costs significantly.

Trust and security are important concerns: adversarial prompt injection attacks, where malicious content in the environment causes an agent to behave in unintended ways, are a known risk in multi-agent deployments. Governance frameworks must address how authorisation, audit logging, and human escalation are implemented across agent boundaries.

References

1. Wooldridge, M., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. The Knowledge Engineering Review, 10(2), 115–152.
2. Wu, Q., Bansal, G., Zhang, J., et al. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. arXiv:2308.08155.
3. MIT Sloan Management Review & Boston Consulting Group. (2025). AI agents: From experimentation to enterprise. MIT SMR/BCG Survey.
4. Dimensions Market Research. (2025). Multi-agent systems market report: 2024–2034. Dimensions MR.
5. Microsoft. (2025, December 4). Single agents to AI teams: The rise of multi-agentic systems. Microsoft Cloud Blog.
6. Google Cloud. (2025). What is a multi-agent system in AI?. Google Cloud Discover. https://cloud.google.com/discover/what-is-a-multi-agent-system