[Webinaire de démonstration] Prêt(e) à dire au revoir à ZooKeeper ? Rencontrez KRaft ! | Inscrivez-vous dès maintenant
Artificial intelligence (AI) agents are everywhere.
Salesforce has Agentforce, Google launched Agentspace, and Snowflake recently announced Cortex Agents. But there’s a problem: They don’t talk to each other.
Your customer relationship management (CRM) agent doesn’t know what insights your data warehouse agent has. Your knowledge retrieval agent operates in isolation. Instead of having a connected AI ecosystem, we’re repeating history and creating AI silos.
Just as software-as-a-service (SaaS) sprawl once trapped enterprise data in disconnected apps, AI sprawl is keeping intelligence locked inside separate platforms. The result? Duplicated work, lost insights, and inefficiencies.
To solve this, we need a shared language for AI agents.
AI agents perceive, reason, and act based on data.
Unlike traditional automation, which follows fixed rules, AI agents adapt dynamically, making real-time decisions based on evolving inputs. They don’t just execute workflows; they continuously refine and optimize them.
Why enterprises need AI agents:
Reduce Information Overload. Employees spend nearly 20% of their week searching for data (McKinsey). AI agents surface insights instantly, improving productivity.
Drive Efficiency and Scalability. Businesses can scale processes efficiently without ballooning costs. AI agents automate complex workflows and adapt to demand without requiring additional headcount.
Enhance Customer Engagement. Agents personalize interactions using real-time insights, automating responses while improving satisfaction.
Accelerate Innovation. By handling routine analysis, agents free up teams for higher-value work.
Leading companies are already leveraging agents to streamline operations. ADT, for example, is using an AI agent to help its customers select, order, and set up their home security.
Vendors are responding with AI agent platforms—Salesforce Agentforce, Google Agentspace, Snowflake Cortex Agents, Glean Agents—that are designed to help enterprises act on their data. But while these tools are powerful, they operate independently, creating AI silos rather than a connected intelligence layer.
AI agents are designed to automate tasks, surface insights, and streamline decision-making. But what happens when these agents operate in isolation?
Imagine a company fully invested in AI agents:
CRM agents manage customer interactions and sales.
Data warehouse agents analyze trends and forecasting.
Knowledge management agents retrieve internal documents.
Custom-built agents automate development tasks like PR reviews.
Each agent excels in its domain but operates in isolation, blocking collaboration.
For example, let’s say a sales rep in their CRM wants to close a deal with a key customer. The rep asks a CRM agent for insights, and it pulls up recent interactions, support tickets, and contract history. But what it doesn’t know is that the data warehouse AI just flagged a new buying trend in the customer’s industry, which could influence the deal. The rep never sees this insight because the warehouse agent and the CRM agent don’t share information.
AI-driven tools should create seamless automation—but without interoperability, they create fragmented intelligence silos, duplicating effort and blocking efficiency.
For years, enterprises have struggled with data silos, where valuable information is locked in separate systems. The industry’s response has been to consolidate data into warehouses and lakes.
With agents, the problem is repeating itself but manifesting in a more insidious way.
Each agent generates intelligence—but without a shared language, those insights remain locked in isolated AI systems, forcing enterprises to manually bridge the gaps.
The solution to agent silos is not the same as the solution to data silos. While consolidating data into a single repository may help with analytics, agents are fundamentally different. They don’t just store knowledge; they act on it. What enterprises need is not just a central hub of information but a way for AI agents to communicate in real time, regardless of where they run.
Large enterprises already struggle with teams, departments, and business units operating in isolation. Sales, marketing, customer support, and finance often have misaligned goals, redundant processes, and disconnected workflows, leading to inefficiencies and wasted effort.
Now agents are replicating this dysfunction at the technology level. Each system’s agent automates tasks within its own domain—CRM agents focus on sales, warehouse agents analyze data, and document retrieval agents surface knowledge. But they don’t coordinate across the enterprise, meaning insights and actions remain locked within individual platforms.
Agents aren’t just software add-ons. They’re decision-makers that must collaborate, exchange insights, and orchestrate actions across the enterprise. Treating them as isolated tools misses the point.
The solution isn’t to avoid agents. It’s to connect them.
To solve the AI silo problem, enterprises need a shared communication layer, not just a central storage solution. Just as APIs enabled SaaS applications to interoperate, we need an equivalent for AI agents—a real-time, event-driven approach that lets agents share intelligence and take coordinated action.
Next we’ll explore how data streaming can provide this missing foundation, allowing AI agents to work together rather than in isolation.
The problem with AI silos isn’t just that agents don’t share data. It’s that they lack a universal way to communicate. Each agent operates within its own platform, following vendor-specific rules without a common language to exchange insights or coordinate actions.
This is where event-driven architectures come in.
Just as data streaming platforms (DSPs) have enabled real-time enterprise data flow by breaking down traditional data silos, they can do the same for AI agents. By providing a shared event stream, DSPs allow agents to act on the same information in real time, ensuring that insights are not just stored but actively shared across AI systems.
A data streaming platform serves as an event-driven backbone for inter-agent product communication.
Instead of relying on point-to-point integrations that create a complex web of NxM dependencies, agents can subscribe to shared event streams, enabling real-time updates, dynamic reactions to changes, and seamless cross-system coordination. This shifts the dependency model to a more scalable N+M approach, where each agent connects to a unified event stream rather than managing direct connections to every other system.
Apache Kafka®, Apache Flink®, and similar technologies already allow enterprises to handle real-time data movement at scale. These same architectures can serve as a communication layer for AI agents, ensuring that they can:
Subscribe to relevant events across different systems, receiving continuous updates as new insights emerge.
Process and react to data in real time instead of operating in isolation with outdated or incomplete context.
Coordinate actions across multiple AI agents, ensuring a seamless flow of intelligence between systems rather than relying on manual intervention.
Enforce governance for data quality and access control, ensuring that inter-agent communication adheres to enterprise policies.
The CRM agents don’t need to know what’s consuming their events. The software can continue to be decoupled and built by a different organization, but the DSP bridges the gap for communication.
While the DSP provides the foundation, fully realizing this vision requires extending its capabilities. This includes having an agent registry for structured governance, provisioning Kafka topics for each agent for seamless event exchange, and stream processing to intelligently route insights across agents in real time.
Let’s explore how these components come together.
Much like how a schema registry ensures structured data governance and defines data products in modern DSPs, an agent registry would provide a standardized framework for registering and managing AI agents across different platforms.
Each agent registers itself in the agent registry, providing:
Name and description
Supported behaviors and capabilities
Expected input types (structured events it listens for)
Expected output types (structured events it produces)
Endpoint for event consumption
This registry ensures that every agent becomes a well-defined data product, discoverable and consumable by other agents without requiring direct integration.
Once an agent is registered, the DSP automatically provisions a dedicated Kafka topic for it.
When new events are written to this topic, they’re picked up by the agent, which can process and act on them.
When an agent generates an output, it writes the result to a default event topic rather than sending it to a predefined endpoint.
This removes tight coupling between agents, allowing outputs to flow dynamically rather than through preconfigured integrations.
With AI agents publishing and consuming events dynamically, we need a way to ensure that relevant outputs reach the right agents without explicit point-to-point mappings.
This is where stream processing (e.g., Flink) and a large language model (LLM)-based event mapper come in:
Stream processing listens to the default event topic, where all agent outputs are published.
It parses the event, determines its relevance, and routes it to the appropriate agent topics based on context.
An embedded LLM can be used to intelligently map outputs to agents, reducing manual configuration and enabling flexible, intent-driven automation.
With this model, agents function much like tools within an agent framework. Use is dynamic and based on supported behavior rather than explicit dependencies. An agent can seamlessly consume another agent’s insights, whether it’s from a CRM, data warehouse, or knowledge retrieval system, without requiring a direct connection.
Some modern agent frameworks, such as Microsoft's AutoGen, employ an event-driven model for message passing, where agents communicate by publishing and subscribing to topics. In AutoGen, each agent’s subscriptions are predefined, allowing for static mapping between topics and agents. This works well in a controlled environment where all agents are built using AutoGen, as their interactions can be anticipated and hardcoded.
However, in a heterogeneous enterprise landscape with agents spanning multiple platforms and technologies, such static mappings break down. Agents are developed independently, using different frameworks, APIs, and data structures, making it impractical to hardwire their interactions in advance.
Instead we need a more flexible and dynamic approach where:
Agents register their capabilities on the fly rather than requiring predefined subscriptions.
Events are intelligently routed using stream processing rather than relying on rigid configurations.
Multi-agent collaboration happens dynamically, regardless of platform or vendor.
By integrating an agent registry with a data streaming platform, agents can self-register, publish events, and dynamically subscribe to relevant event streams without requiring explicit integrations or vendor lock-in. This enables a truly open, real-time AI ecosystem, where intelligence flows freely across platforms.
For agents to deliver on their full potential, interoperability must be a priority.
Agents can no longer operate as isolated systems requiring custom integrations for every interaction. Instead they must be part of an open, event-driven ecosystem where intelligence flows seamlessly across platforms.
By treating agents as event-driven, interoperable data products, enterprises can:
Enable real-time collaboration across AI platforms without rigid integrations.
Avoid vendor lock-in by using shared event streams rather than proprietary APIs.
Scale efficiently, as each agent only needs to register and consume relevant events instead of managing direct connections.
Just as data streaming solved the enterprise data silo problem, it now provides the foundation for breaking down AI silos. The future of AI isn’t a collection of disconnected tools. It’s an interconnected network of agents that communicate, share insights, and act together in real time, using event-driven data streaming as their shared language.
Apache®, Apache Kafka®, Kafka®, Apache Flink®, and Flink® are registered trademarks of the Apache Software Foundation in the United States and/or other countries. No endorsement by the Apache Software Foundation is implied by using these marks. All other trademarks are the property of their respective owners.
Introduction to Flink SQL FEDERATED_SEARCH() on Confluent cloud. FEDERATED_SEARCH() along with ML_PREDICT() enables developers to execute GenAI use cases with data streaming technologies.
Learn how Flink enables developers to connect real-time data to external models through remote inference, enabling seamless coordination between data processing and AI/ML workflows.