Multi-agent AI systems, where several specialized AI agents work together on complex problems, are failing at alarming rates because of design flaws rather than technological limitations. A new analysis of over 1,600 execution traces from seven popular frameworks reveals that poor task specification, miscommunication between agents, and weak verification processes are the primary culprits, not the underlying AI models themselves.

What Are Multi-Agent AI Systems and Why Do They Matter?

A single AI agent can handle straightforward tasks like summarizing a contract or answering a refund question. But when faced with complex work like reviewing 400 supplier contracts, flagging risky clauses, checking them against procurement policy, and drafting a report for the CFO, a single agent starts to slip. It loses track of earlier details, skips steps, and sometimes marks its own homework.

Multi-agent systems solve this by splitting the work across specialized agents, each with a clear role and its own tools. One agent might pull data from your customer relationship management (CRM) system, another analyzes it, a third writes the summary, and a fourth checks that summary against the source numbers before anyone sees it. This division of labor means each agent only holds the context for its own piece of the job, preventing the cognitive overload that degrades single-agent performance.

Why Do Multi-Agent Systems Fail So Often?

Researchers at UC Berkeley studied this question in a paper called "Why Do Multi-Agent LLM Systems Fail?" presented at NeurIPS 2025. They collected more than 1,600 annotated execution traces from seven popular multi-agent frameworks and found 14 distinct ways these systems break down. These failures fall into three main categories:

• Specification Problems: The task or the roles were badly defined, so agents misread the goal, step outside their role, or stop before the job is done.
• Misalignment Between Agents: Agents talk past each other, ignore a colleague's input, or lose important details during a hand-off between systems.
• Weak Verification: Nobody checks the output properly, or the checker confirms that something was produced without confirming it is actually correct.

The critical insight is that many of these failures come from how the system is designed, not from a weak underlying model. A smarter language model (LLM) will not fix a vague brief or a missing review step.

How to Design Multi-Agent Systems That Actually Work

• Clear Role Definition: Each agent's job must be narrow enough to describe in a paragraph and to check against a clear standard. If you cannot describe the job that clearly, splitting it across agents usually makes things worse, not better.
• Appropriate Task Structure: Ask whether the work splits into parts that can run independently. If yes, parallel agents can save real time. If every step depends on the one before it, a single agent with good tools is often the better choice.
• Robust Verification Processes: Build in explicit checking steps where one agent reviews another's work before it moves forward. Banks have used the maker-checker rule for decades for the same reason: the person who does the work should not be the one who approves it.
• Cost-Benefit Analysis: Determine whether the result is worth the extra cost. Multi-agent systems burn many times more computing resources than single-agent approaches, so the performance gain must justify the expense.

What Does Success Look Like?

The clearest public example comes from Anthropic, which described the setup behind Claude's Research feature in June 2025. A lead agent plans the research and starts several sub-agents that search in parallel, each with its own context window, then combines what they find. On Anthropic's internal research evaluation, this setup, with Claude Opus 4 leading and Claude Sonnet 4 sub-agents doing the searching, beat a single Claude Opus 4 agent by 90.2%. It performed best on breadth-first questions, where the answer requires looking in many directions at once, such as finding the board members of every IT company in the S&P 500.

However, the cost is substantial. Multi-agent systems used roughly 15 times more tokens (the units of text that AI models process) than a normal chat. Token usage alone explained about 80% of the variation in performance on Anthropic's browsing test, meaning a good share of the gain comes from spending more computing power on the problem.

Where Multi-Agent Systems Deliver Real Business Value

The approach works best where a job already involves several specialists passing work to each other. Real-world applications include customer support, where a triage agent reads the ticket and routes it to billing, policy, or drafting agents; finance operations, where one agent extracts invoice data, another matches it against purchase orders, and a third flags mismatches; market and competitor research, where an orchestrator sends workers to cover pricing, product changes, hiring, and news for each competitor; software delivery, where separate agents write code, tests, and documentation; and compliance review, where one agent pulls clauses from contracts, another compares them with internal rules, and a checker confirms every flagged item points to a real clause.

The Growing Risk of Uncontrolled Agent Deployment

Beyond design failures, organizations face mounting pressure to control what agents can access and do. NVIDIA OpenShell 0.1.0, an open-source runtime released recently, provides a way to enforce which systems and data an AI agent can access without rewriting the agent itself. The runtime combines sandboxed execution, controlled service access, credential management, and formal policy analysis to restrict API operations and protect credentials outside the agent workload.

OpenShell uses three core components to manage agent fleets. The OpenShell Gateway manages the lifecycles and policies of many sandboxes. The OpenShell Supervisor, paired with each sandbox, runs outside the agent workload and checks outbound requests against policy. The OpenShell Sandbox runs the workload with kernel-level controls over its filesystem and processes, with no network path except through the supervisor.

Organizations including Cadence, Slack, and Gecko Robotics are adopting OpenShell for chip design, enterprise automation, and physical robotics governance. The tool allows teams to grant agents the capabilities a task requires while enforcing those permissions outside the workload, addressing a critical gap in agent governance.

The Bottom Line for Organizations

In June 2025, Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value, or poor risk controls. Multi-agent systems, with bigger compute bills and more moving parts, are exposed to all three risks.

Before building a multi-agent system, organizations should ask blunt questions about whether the work truly splits into independent parts, whether the result justifies the extra cost, and whether they have the governance infrastructure to control what agents can access. The technology is powerful, but success depends far more on thoughtful design and oversight than on having the smartest underlying model.