What are multi-agent systems?

A multi-agent system is a group of LLM agents that work together on one task, each running its own loop of model calls and tool use. Anthropic defines it as “multiple agents (LLMs autonomously using tools in a loop) working together.” It can beat a single agent on broad tasks, but it costs far more tokens and adds failure modes of its own.

New to agents? Start with what are AI agents. For the control patterns (sequential, handoff, group chat), see agent orchestration. Here the focus is the part around them: when several agents earn their cost and how they break.

What it is

Common definitions agree on the shape. IBM calls it AI agents “working collectively to perform tasks” and lists flexibility, scalability and domain specialization among the benefits. Anthropic’s research system uses an orchestrator-worker design: a lead agent coordinates the work and hands pieces to subagents that run in parallel, each with its own context window (Anthropic, June 2025).

The unit that matters is the agent loop from agentic workflows and ReAct agents. A multi-agent system runs several of them and passes information between them.

Why more agents can help

Anthropic’s explanation is blunt: “Multi-agent systems work mainly because they help spend enough tokens to solve the problem.” On its BrowseComp evaluation, three factors explained 95% of performance variance, and token usage alone explained 80%. Subagents add capacity by working in parallel and by compressing what they find before reporting to the lead agent.

In Anthropic’s June 2025 test, a Claude Opus 4 lead with Claude Sonnet 4 subagents “outperformed single-agent Claude Opus 4 by 90.2%” on its internal research eval. That is one vendor’s result on one kind of task, not a general law.

What it costs

Setup	Token use vs a chat	Source
Single agent (Anthropic, June 2025)	about 4 times more	Anthropic
Multi-agent system (Anthropic, June 2025)	about 15 times more	Anthropic

Anthropic’s early agents also showed waste, in its words: “spawning 50 subagents for simple queries,” “scouring the web endlessly for nonexistent sources,” and “distracting each other with excessive updates”.

Where it fits poorly

Anthropic says domains that need all agents to share the same context, or have many dependencies between agents, are a poor fit today. Most coding tasks, it notes, have fewer parallelizable parts than research (Anthropic).

Cognition made the strongest case against. Walden Yan wrote in June 2025 that running multiple agents “only results in fragile systems,” because decisions get dispersed and context is not shared thoroughly enough. Two principles follow: “Share context, and share full agent traces, not just individual messages,” and “Actions carry implicit decisions, and conflicting decisions carry bad results” (Cognition).

Cognition has since narrowed that view. In April 2026 Yan wrote that “a narrower class of patterns” works: “setups where multiple agents contribute intelligence to a task while writes stay single-threaded.” He gives a clean-context reviewer that catches bugs the coder cannot see as one example (Cognition). Anthropic’s and Cognition’s accounts agree on the deciding factor, which is shared context.

How they fail

The paper Why Do Multi-Agent LLM Systems Fail? (first posted 17 March 2025, revised 26 October 2025) built a taxonomy called MAST from 150 traces, with annotator agreement of kappa 0.88. It names 14 failure modes in three categories, and its dataset has over 1,600 annotated traces from 7 frameworks. The authors say multi-agent systems often show only minimal gains on popular benchmarks, which motivated the study.

Category	Examples (prevalence reported in the paper)
System design issues	Step repetition (15.7%), unaware of termination conditions (12.4%), disobeying the task specification (11.8%)
Inter-agent misalignment	Reasoning-action mismatch (13.2%), task derailment (7.4%), failing to ask for clarification (6.8%)
Task verification	Incorrect verification (9.1%), no or incomplete verification (8.2%), premature termination (6.2%)

Prevalence figures are from section 4 of the paper. IBM adds that agents built on the same foundation model can share the same weaknesses, and that decentralized agents can behave in conflicting or unpredictable ways.

Deciding

Use the evidence above as a checklist:

• The task splits into parts that can run in parallel and need little shared context: multiple agents can pay off.
• Steps depend on each other or on one shared picture of the work: one agent is the safer start.
• Whatever you build, give agents full context, stop conditions and a verification step, since those are where the MAST failures cluster.

How to build AI agents covers starting with one agent. For agents that talk across systems, see the A2A protocol.

Tracing a multi-agent run

Failures like step repetition and derailment show up only when each agent’s calls sit in one trace; see agent observability. HoneyHive traces LangGraph graphs, node transitions and LLM calls, with setup in the HoneyHive LangGraph integration docs.

FAQ

What is a multi-agent system?

Anthropic defines it as "multiple agents (LLMs autonomously using tools in a loop) working together" (Anthropic). IBM describes it as AI agents "working collectively to perform tasks".

Are multi-agent systems better than a single agent?

Sometimes. Anthropic reported a multi-agent research setup beat a single agent by 90.2% on its internal research eval, but said the approach works mainly because it spends more tokens. Cognition argued in June 2025 that multiple agents in collaboration "only results in fragile systems", then in April 2026 said it runs setups where "multiple agents contribute intelligence to a task while writes stay single-threaded" (Cognition).

Why do multi-agent systems fail?

A 2025 study of annotated traces from 7 frameworks grouped 14 failure modes into system design issues, inter-agent misalignment and task verification (arXiv).

What is the difference between a multi-agent system and agent orchestration?

A multi-agent system is the set of agents. Orchestration is how you control which one runs next. See agent orchestration for the patterns.