Multi-agent AI promises human-capable intelligence and reasoning with machine-capable speed and scale. However, today’s multi-agent systems fail frequently and in predictable ways. New research shows the root cause isn’t a lack of intelligence, but a lack of systemic visibility, coordination, and verification.

Key findings

• Multi-agent AI systems fail at high rates, often exceeding 50% across production workloads

• Stronger models will not correct these failures as they are systemic, not model-specific

• Most breakdowns stem from poor system design, coordination gaps, and weak verification

• Adding more agents often increases failure risk without better observability

• Reliable multi-agent systems require continuous analytics, not post-hoc review

The causes behind multi-agent system failures

In theory, multi-agent systems (MAS) are designed to mimic teams. They’re expected to divide work, reason in parallel, and, ultimately, converge on better outcomes. The reasoning from overconfident executives and overpromising vendors is that these systems should outperform single agents. In practice, they often don’t.

Large-scale analysis of real multi-agent executions shows that failure isn’t random. The same failure patterns repeat across frameworks, models, and task types. What’s interesting about these failures isn’t the agents. In isolation, the agents are capable. Where things fall apart is due to an inherent inability to observe MAS as they execute and course correct while they’re operating. 

Based on research from UC Berkeley and Intesa Sanpaolo, MAS failures can be grouped into three major domains: system design, agent coordination, and poor verification.

System design breaks before reasoning begins

Many failures originate upstream, well before agents generate answers. Tasks are underspecified or vague. Roles may overlap or conflict with each other. Stop conditions aren’t clear. Agents get trapped in loops, redo work, or proceed without knowing when the job is done.

These aren’t errors in the models, or even shortcomings of mathematical reasoning. These errors are flaws in how MAS are designed, observed, and controlled. Agent responsibility isn’t clearly defined and limited. Coupled with the inability to continuously monitor progress, it’s no wonder MAS, even with highly capable individual agents, go off the rails more often than not. 

Agents don’t coordinate

Much like SecOps and ITOps teams must coordinate internally, the same is true in MAS. Despite breathless vendor promises, agents aren’t magic. They need the correct information at the right time, and that information comes from other agents. In real world scenarios, agents are often starved for data from peer agents.

This occurs because agents are horrible co-workers. They hoard critical details or misinterpret input, or even outright ignore it. Lacking the correct data, they’ll make incorrect assumptions because there’s usually no one to ask for clarification. Even better, sometimes they “agree” in conversation but execute something else entirely. 

We’ve all worked with those people. Do we really need agents copying passive-aggressive behavior? 

What’s really missing here is situational awareness. There’s no unified view of intent, state, and outcomes between agents. It’s no wonder that coordination collapses. 

Verification is thin if it completes at all

Many MAS perform some form of validation, but it’s usually superficial: a script runs, the response looks reasonable, and a box is checked.

What’s missing is some notion of layered verification. Systems rarely validate whether the result actually meets the original objective, whether intermediate steps were correct, or whether errors compounded across agents. As a result, failures pass unnoticed until the final output is already wrong.

Why this matters for the enterprise

The takeaway is stark for IT and security leaders: multi-agent systems require system-level design and architecture. These aren’t novelties. MAS produces signals, actions, and decisions that must be observed, analyzed, and governed like any other complex system.

The biggest risk isn’t that AI agents make mistakes. Of course they will. The risk is that they make mistakes quietly, without visibility into how or why they happened.

This is where most current implementations fall short and where Cribl’s strengths are directly relevant.

Building more effective multi-agent systems

Reliable multi-agent AI requires the same discipline applied to distributed systems, security operations, and large-scale analytics.

First, treat agent interactions as data, not a conversation. Every decision, handoff, and action is a signal. If those signals are fragmented across tools or buried in logs, systemic issues remain hidden. Centralized analytics across distributed agent activity is foundational.

Second, federate insight, not just execution. Multi-agent systems often span models, tools, and environments. Without the ability to query and correlate activity across those domains in real time, teams lose the ability to reason about system behavior as a whole. Federated data access enables unified analysis without forcing everything into a single platform, data format, or pricing model.

Third, measure progress continuously. Effective systems don’t wait until the end to validate outcomes. They analyze intermediate states, detect drift, and surface anomalies early. This requires flexible, AI-augmented analytics that can adapt as workflows evolve.

Finally, design verification as an analytics problem. Validation should answer business and operational questions, not just technical ones: 

• Did the system meet the objective? 

• Did agents converge or diverge? 

• Where did confidence degrade? 

These answers come from querying data, not trusting intent. 

Wrapping up

Multi-agent systems don't fail because they're too advanced. It fails because it’s treated as magic rather than machinery.

The path forward isn’t more agents or more complexity. It’s better visibility, better analytics, and better feedback loops. Enterprises approaching multi-agent systems with the same rigor they apply to security operations and distributed infrastructure will be the ones that realize real value.

As AI systems grow more autonomous, the ability to observe, analyze, and govern their behavior across environments becomes non-negotiable. This sounds insurmountable but it isn’t. These aren’t AI problems. They’re data problems. And we already know how to solve those.