Enterprises are racing to build AI agents for customer support, procurement, coding, and planning. But the real transformation begins when the agents start working together.
Recent research by HFS Research and Cognizant shows that 73% of surveyed companies are already running multi-agent systems. From Telecom and banking to insurance and travel, enterprises are employing agents for underwriting decisions, network routing, trading optimization, procurement negotiation, and customer-facing workflows.
Companies leading this shift—what our report calls “Orchestrators”—already run an average of 12 agents in production, with some environments reaching as many as 20.
At this point, agents stop behaving like simple tools and start functioning more like digital employees. They coordinate tasks, share context, interact with each other, and occasionally behave in unanticipated ways.
This fundamentally changes how enterprises must think about AI. The moment organizations move from one agent to many, they’re designing ecosystems likely to exhibit emergent behavior.
And those ecosystems require something enterprises largely lack today: an agent operating system. This system is a way to run, manage and orchestrate networks of agents while allowing them to take action using defined tools or applications in a safe and reliable way.
Enterprise AI’s next phase will be defined by how those agents are coordinated, governed, and managed.
Where the problems start
Multi-agent systems can consist of autonomous components that interact in dynamic environments. Our research shows that at around five agents, coordination problems begin to appear when agents create their own workflows or unexpected shortcuts.
The research found that:
- 21% of enterprises experienced cascading failures when a single agent malfunctioned
- 22% encountered emergent behaviors, where agents developed unexpected workflows or decision patterns
These outcomes reflect a new computing paradigm in which organizations must engineer, monitor, and govern networks of autonomous and semi-autonomous systems.
The AI adaptation gap
Another trend emerging from our research is the AI adaptation gap, which also points to the need for a new architectural layer.
We found that 24% of companies moved from proof of concept to production in under three months. The technology evolves rapidly, but enterprise governance models, operational frameworks, and architectures struggle to keep pace.
When that gap widens, organizations begin to encounter problems such as auditability gaps where thousands of AI decisions have no clear lineage and data inconsistencies are amplified across multiple agents.
One of the most surprising outcomes of multi-agent deployments is that agents do more than automate work; they reveal how work actually happens.
Several organizations found that AI deployments exposed hidden inefficiencies, conflicting key performance indicators, and fragmented workflows. One executive described the experience as “an expensive audit of our dysfunction.”
While uncomfortable, this insight can be powerful. Agent systems highlight the gap between how processes are documented and how work really flows through the enterprise. Organizations that invest in accurately mapping workflows and context will be far better positioned to redesign operations for a more autonomous future.
Enterprise needs an agent operating system
If agents behave more like employees than traditional software modules, enterprise organizations need infrastructure that manages them accordingly. This is where the concept of an Agent Operating System (Agent OS) becomes critical. An Agent OS provides the structured layer that governs how agents interact with enterprise data and with human teams.
Our research identifies five foundational layers required to support scalable agent ecosystems:
1. Governance and autonomy management: Define levels of agent autonomy, determine when human oversight is required, and establish escalation mechanisms.
2. Orchestration: Standardize communication protocols and coordination mechanisms to collaborate effectively.
3. Observability and explainability: Ensure visibility into how agent decisions are made and how behaviors evolve over time.
4. Data trust: Lineage, quality monitoring, and validation are essential, because multi-agent systems amplify data inconsistencies.
5. Human–agent workforce management: Rethink collaboration models between humans and AI, and define accountability, roles, and performance metrics for mixed teams.
Together, these layers form the foundation for scalable, trustworthy agent ecosystems.
Multi-agent architectures can mediate between systems, translate context, and generate consolidated insights across distributed environments. This approach allows enterprises to extract value from fragmented data landscapes while longer-term data modernization efforts continue.
Design the future deliberately
To succeed in the next phase of enterprise AI, organizations must rethink how they design and scale agentic systems. Three principles stand out:
- The real value of AI lies in networks of collaborating agents, not isolated copilots.
- Trust, governance, and observability must be embedded from the start, not added after deployment.
- Agent Operating Systems offer a structured layer that orchestrates agents, governs autonomy, ensures data integrity, and manages human–AI collaboration.
Enterprises that build this foundation will move beyond isolated AI experiments to scalable, enterprise-wide intelligent automation. And ultimately, those organizations will define the future of enterprise AI. Enterprise agentification is like peeling an onion: the more agentified, the more the need to agentify. A solid Agent Operating System can eliminate the tears as you peel.
Babak Hodjat is the chief AI officer at Cognizant.