All postsAI explained

Multi-agent systems: why more agents often fail

APJ-RAISE team · · 7 min read

A multi-agent system is a group of AI agents that share one task, each with its own role, tools and instructions, passing work to each other until the job is done.

Close-up of red ants working together on dark soil
Photo: Глеб Коровко, Pexels

The idea sounds like teamwork, so more agents should mean better results. Often they do not. This post asks one question: why do more agents so often make a system worse? The answer comes from simple counting, one large failure study and the cost numbers that vendors publish.

If you are new to agents, read what AI agents are first. An agent here is a language model that uses tools in a loop: it picks a step, acts, checks the result and repeats.

What is a multi-agent system in AI?

It is software where two or more agents split a goal. A typical research assistant has a planner that breaks your question into parts, two searchers that work in parallel, and a writer that merges their notes.

The field is older than chatbots. Michael Wooldridge's textbook An Introduction to MultiAgent Systems (2nd edition, Wiley, 2009) covers agents that negotiate, vote and bid in auctions. Today's LLM agents borrow much of that thinking.

Four layouts cover most systems you will meet:

  • Orchestrator-worker: a lead agent plans and hands parts to workers, then joins the results. Claude's Research feature works this way.
  • Pipeline: agents act in a fixed order, like a factory line. Easy to test, but one bad step spoils the rest.
  • Debate: several agents answer, read each other's answers and revise. Yilun Du and colleagues used this in 2023 to improve factual accuracy and reasoning.
  • Blackboard: agents share one workspace instead of messaging. The idea goes back to the Hearsay-II speech system, described in 1980.

Why do multi-agent systems get harder to test as they grow?

Every pair of agents is a link where a message can be misread. With n agents there are n(n-1)/2 possible pairs. The number of agents grows in a straight line, but the pairs grow much faster. Our table:

  • 2 agents: 1 pair
  • 4 agents: 6 pairs
  • 8 agents: 28 pairs
  • 15 agents: 105 pairs

Doubling from 4 to 8 agents multiplies the pairs by almost five, from 6 to 28. The 105 figure for 15 agents is the one Capgemini used on 17 November 2026, when it named multi-agent systems the first of its five tech trends for 2027. It added that 1,000 agents and 200 agentic systems give almost 720,000 pairs. Capgemini did not show the working. Our reading: treat them as 1,200 units, and 1,200 × 1,199 ÷ 2 = 719,400.

Capgemini summed up the risk neatly: each agent can be right while the whole system is wrong.

How small errors add up in an agent pipeline

Pairs are only half the story. In a pipeline, each step must go right for the final answer to be right. The next example is made up for illustration, not a measured result. Suppose each agent gets its step right 95% of the time, and the errors are independent:

  1. 1With 2 agents: 0.95 × 0.95 = about 90% of runs are fully correct.
  2. 2With 4 agents: 0.95 to the power 4 = about 81%.
  3. 3With 8 agents: 0.95 to the power 8 = about 66%.

So a chain of eight good agents fails about one run in three. Real systems are messier, because errors are rarely independent and a later agent sometimes catches an earlier mistake. The point stands: every agent you add needs a reason to be there and a way to check its output.

Why do multi-agent LLM systems fail? What MAST found

The best evidence is the paper Why Do Multi-Agent LLM Systems Fail? by Mert Cemri and 12 co-authors, posted on arXiv in March 2025 and revised in October 2025. The team annotated over 1,600 traces (step-by-step logs of runs) from 7 frameworks and built a taxonomy called MAST, with 14 failure modes in 3 groups.

The first group is system design: wrong roles, a bad task setup or missing stop rules. The second is inter-agent misalignment, where agents misread, ignore or contradict each other. The third is weak verification, when nobody checks the result or the check is wrong.

Notice that two of the three groups are about structure, not about how clever each model is. A smarter model does not fix a team with no stop rule. The authors also report that gains over single agents on popular benchmarks are often minimal.

Two more failure types grow with team size:

  • Prompt injection relays: hidden text in a file gives a worker new orders, and the worker passes them up. In October 2026, BleepingComputer reported that a malicious AGENTS.md file in a Downloads folder could hijack AI coding agents.
  • Runaway cost: Anthropic lists spawning 50 subagents for simple queries among the early failures of its own system.

Multi-agent vs single agent: when is a team worth it?

The strongest pro-team number comes from Anthropic. In June 2025 it reported that a Claude Opus 4 lead with Claude Sonnet 4 workers beat a single Claude Opus 4 agent by 90.2% on its internal research test.

Most coverage stopped there. The same post says multi-agent runs used about 15 times more tokens than chats, and that token use alone explained 80% of score differences on a browsing test. Read together, much of the gain is extra compute spent in parallel. These are Anthropic's figures on Anthropic's tests, so they are not a general law.

Our rule of thumb from the evidence above:

  • Use one agent when the task is a single chain of reasoning, or when the budget is tight.
  • Use a team when the work splits into parts that do not depend on each other, such as searching many sources at once.
  • Add a checker before you add another worker. MAST shows verification is where many systems break.
  • Keep a list of every agent, its permissions and a way to stop it. Capgemini asks for the same, and Gartner lists AI agent identity among its 2027 trends.

How do AI agents talk to each other? MCP and A2A

Two open protocols carry most agent traffic in 2026. MCP, the Model Context Protocol from Anthropic, connects an agent to tools and data. Its current specification, version 2026-07-28, made it stateless: each request carries its own version information, with no opening handshake.

A2A, the Agent2Agent protocol, lets agents from different vendors find each other and hand over tasks. On 9 April 2026 the Linux Foundation said more than 150 organisations backed it, that version 1.0 was its first stable specification, and that Google, Microsoft and AWS had built it into their clouds. MCP, Block's goose and OpenAI's AGENTS.md became founding projects of the Linux Foundation's Agentic AI Foundation on 9 December 2025.

A short way to remember it: MCP is an agent picking up a tool, A2A is an agent calling a colleague.

Test the "more agents" problem yourself

You can run all of these on a laptop with LangGraph, CrewAI or the OpenAI Agents SDK, all MIT licensed and active on GitHub in November 2026. Avoid Microsoft's AutoGen for new work; it is archived and gets no more fixes.

  1. 1Plant one wrong fact in the first agent of a pipeline. Measure how often it reaches the final answer with 2, 4 and 8 agents, and compare with our 95% illustration.
  2. 2On 100 school-level maths questions, compare one agent with three debating agents. Report accuracy and token cost side by side.
  3. 3Save 50 traces from a three-agent system and label each failure with MAST. The MAST data are public, so you can compare your labels with theirs.

If you use an AI model to label traces, read our guide to LLM-as-a-judge and check a sample by hand. For a real agent funnel at scale, see how OpenAI's agents found Linux kernel exploits. To build one with other students, join APJ-RAISE.

Questions people ask

Why do multi-agent LLM systems fail?
The MAST study of over 1,600 traces found three groups of causes: poor system design, agents misreading each other, and weak checking of the final result.
Is multi-agent better than single agent?
Not always. Anthropic saw a large gain on its research test but used about 15 times more tokens than chats. A team helps when the work splits into independent parts.
What is the difference between MCP and A2A?
MCP connects an agent to tools and data. A2A lets agents talk to and hand tasks to other agents. Both are open protocols hosted by the Linux Foundation.

Sources

  1. 1.Capgemini. TechnoVision: Top 5 Tech Trends to Watch in 2027. Press release, Paris, 17 November 2026.
  2. 2.Cemri, M., Pan, M. Z., Yang, S. and others. Why Do Multi-Agent LLM Systems Fail? arXiv:2503.13657, 2025.
  3. 3.Anthropic. How we built our multi-agent research system. 13 June 2025.
  4. 4.Du, Y., Li, S., Torralba, A., Tenenbaum, J. B. and Mordatch, I. Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.14325, 2023.
  5. 5.Wooldridge, M. An Introduction to MultiAgent Systems, Second Edition. John Wiley & Sons, 2009.
  6. 6.Erman, L. D., Hayes-Roth, F., Lesser, V. R. and Reddy, D. R. The Hearsay-II Speech-Understanding System: Integrating Knowledge to Resolve Uncertainty. ACM Computing Surveys 12(2), 1980.
  7. 7.Model Context Protocol. Versioning (specification documentation).
  8. 8.Linux Foundation. A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms, and Sees Enterprise Production Use in First Year. 9 April 2026.
  9. 9.Linux Foundation. Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md. 9 December 2025.
  10. 10.Toulas, B. AI coding agents hijacked using AGENTS.md file in Downloads folder. BleepingComputer, 25 October 2026.
  11. 11.LangGraph on GitHub (langchain-ai/langgraph).
  12. 12.CrewAI on GitHub (crewAIInc/crewAI).
  13. 13.OpenAI Agents SDK on GitHub (openai/openai-agents-python).
All posts