A general AI model applied to a complex workflow tends to lose coherence as the number of steps increases. Instructions contradict each other, context is lost, and intermediate results are not transferred correctly between phases. Multi-agent systems were created to solve this problem: instead of using a single model for everything, the workflow is divided into steps handled by specialized agents, each optimized for the specific task it must perform. Coordination between agents becomes the problem to solve, but the result is more robust and scalable automation.
Interest grew 1,445% in just over a year
Gartner recorded a 1,445% growth in inquiries about multi-agent systems between Q1 2024 and Q2 2025. Interest is so high that Gartner has included multi-agent systems among the top strategic technology trends for 2026, defining them as collections of AI agents that interact to achieve complex individual or shared goals. Gartner predicts that by 2027, 70% of multi-agent systems will use tightly specialized agents, increasing accuracy but also increasing coordination complexity.
The main advantage over a single agent is modularity. A specialized agent for reading documents, alongside an agent that validates extracted data, one that queries an ERP system, and one that produces the final report, delivers more reliable results than any general model trying to do everything on its own. Proven agents can be reused across different workflows, increasing scalability and reducing the errors that typically afflict monolithic AI systems. Multi-agent systems also allow different organizations to connect their agents, paving the way for what Gartner calls the "Internet of Agents."
Risks that grow with complexity
The very feature that makes multi-agent systems powerful, namely the multiplication of interactions between autonomous components, is also the source of their main risks. Gartner identifies four categories of criticality. The first is a larger attack surface: each agent is a potential vulnerability point, and agents can be manipulated through prompt injection, poisoned data, or fraudulent websites they interact with during task execution. The second is error propagation: an error produced by one agent amplifies as it is passed to subsequent agents, producing overall outputs far more distorted than the original error. The third is integration and monitoring complexity: more agents mean more connections, more logs, more surface to observe. The fourth is cost management, which can become unpredictable when systems operate autonomously at high volumes.
Added to these risks are specific threats flagged by Gartner: credential hijacking, interactions with untrusted sources, and deviant agent behavior caused by internal flaws or external triggers. By 2028, according to Gartner, 40% of CIOs will require that "guardian agents" be available, agents specifically designed to monitor, control, and contain the actions of other agents.
How to design a multi-agent system that works
Gartner outlines four design principles for enterprise multi-agent systems. The first is a modular event-driven architecture with semantic layers: each agent must be independent, communicate through events rather than direct calls, and access data through a common semantic layer that ensures consistency and security. The second is governance and observability from the start: do not add governance as a later layer, but embed it in the design from the first version. Real-time audit trails, action monitoring, and compliance checks must be integral parts of the architecture. The third is the use of orchestrated workflows with output validation at every step: the system should not blindly execute a chain of actions but validate intermediate results before proceeding. The fourth is the adoption of open, interoperable standards for agent-to-agent communication, to avoid vendor lock-in and enable integration of third-party agents as the market matures.
Gartner explicitly recommends starting with small pilots equipped with observability tools before scaling. The problem with multi-agent systems is not the demonstration phase, but production behavior at high volume. High-impact but low-complexity use cases, such as structured document management or responding to standard back office requests, are the most appropriate starting point.
Gartner estimates about 130 real vendors out of thousands claiming to be agentic
The multi-agent and agentic systems market suffers from the same problem that plagues agentic AI in general: a very high number of vendors using the term without real underlying capabilities. Gartner estimates that out of thousands of vendors calling themselves "agentic," about 130 have truly agentic capabilities. For multi-agent systems, evaluation must go beyond vendor claims and verify shared memory management between agents, rollback capabilities in case of errors, handoff mechanisms between agents, and security guarantees for access to data and external systems.
The end state Gartner outlines is not a system that runs on its own without oversight, but a system that drastically reduces human intervention in standard cases, manages exception propagation transparently, and allows operators to understand what it is doing and why. Trust in agents is built through observability, not through blind autonomy.
