Confluent Unveils AI Streaming Agents Powered by Kafka Data
Confluent has unveiled its new Streaming Agents feature, now available in open preview within Confluent Cloud. This new functionality marks a significant evolution in the application of artificial intelligence, enabling organizations to select models, engineer prompts, specify tools and data sources, implement testing, and enrich data for event-triggered, multi-agent AI systems.
Confluent’s approach to AI agents is notable for two compelling reasons. Firstly, it aims to fully realize the promise of smart agents by feeding them the most current data directly from Confluent’s Apache Kafka foundation, thereby equipping them with up-to-the-minute situational awareness. Secondly, and perhaps more profoundly, it flips the conventional AI interaction model. Instead of humans initiating AI agents, typically through natural language prompts to chatbots, Streaming Agents are designed as autonomous entities that react dynamically to low-latency enterprise data. Sean Falconer, Confluent’s head of AI, explains that Streaming Agents are partly conceived to propel businesses toward a future where AI is characterized by “ambient, event-driven agents that are embedded in your infrastructure, monitoring the state of the business, and reacting based on the changes of that state.” These AI-powered agents, he suggests, can serve as the “eyes” and “ears” for organizations, gleaning the latest data-driven developments impacting business cases and ensuring intelligent systems adapt immediately. By coupling Kafka’s robust storage and messaging capabilities with the analytics, AI models, and Streaming Agents found in Confluent Cloud for Apache Flink, Confluent’s new feature has the potential to achieve precisely this.
The development and deployment of these Streaming Agents occur within Confluent Cloud for Apache Flink. Flink serves as the computational backbone, enabling agents to seamlessly interact with various external tools and resources. Users can connect to emerging standards like MCP (a protocol for agents to call out to resources), alongside vector and SQL databases, and API endpoints, all managed securely within the Flink environment, according to Falconer. Flink also provides the interface for organizations to define the specific actions they want agents to undertake and to integrate a diverse array of AI models, including those from industry leaders like Gemini, OpenAI, and Anthropic. Falconer notes that the choice of model often marks the initial step in an agent’s construction.
A crucial aspect of agent configuration in Confluent Cloud for Apache Flink involves defining prompts, a process central to “context engineering.” These prompts are twofold: system prompts, which establish the agent’s overarching role within a larger workflow, and task-specific prompts, which detail the precise job characteristics the agent is expected to perform. For instance, a system prompt might designate an agent as “an expert in writing emails,” followed by explicit instructions to “write an email based on the following input that describes this particular user.” This delineation narrows the agent’s scope, clearly codifying its expected inputs and outputs. As Falconer illustrates, “If I’m scoring a lead, then I know the input is going to be a lead and the output should be a score between one and 100.” This precision aids in immediate error detection and facilitates easier testing and evolution of agents.
While Apache Flink provides the processing power, much of Streaming Agents’ inherent power stems from its deep integration with Apache Kafka for inter-agent communication. Rather than relying on alternative protocols, Streaming Agents leverage Kafka’s robust messaging capabilities to enable agents to hand off distinct parts of a complex workflow, collectively achieving larger objectives, such as detecting network failures for a telecommunications provider. This process often begins with diverse, real-time inputs—like weather reports or IoT sensor data—representing the dynamic state of the business. As Falconer elaborates, “Once an agent takes that data, it also talks to some external system to gather some additional context. Then it spins out an output which probably fans out to multiple systems, including other agents, where those agents then operationalize the output.” Typically, these initial events flow into Kafka topics, from where the data can be routed to Flink for aggregation or processing. After an agent acts on this data, it returns its output to Kafka, often to new topics, thereby prompting subsequent agents to continue the workflow. Kafka’s enduring messaging and storage capabilities are invaluable here, providing critical “traceability between agent one and agent two.” This continuous record allows developers to review communication history, facilitating testing against real traffic and enabling the refinement of agent behaviors and models, ensuring improved outputs.
Confluent has also incorporated comprehensive testing capabilities, including “dark launch” support for Streaming Agents. This allows an agent to operate within production traffic without directly interacting with end-users, enabling organizations to deploy a second version—perhaps with a different model or prompts—and measure its performance against the original without impacting live operations. This parallel processing allows for performance comparison, ensuring that “version two is better than the first one.” Users can also conduct A/B testing within Confluent’s environment and enrich agent data by accessing external tables, such as those in MySQL, directly from the streaming data platform.
By augmenting AI agents with real-time streaming data, Confluent is significantly enhancing their capacity to act autonomously on behalf of an enterprise. Falconer aptly compares this real-time intelligence to the “eyes and ears” of a business, complementing insights derived from historical data. He emphasizes that “Most business cases need both”—understanding past customer behavior alongside their current actions. Equipping AI-infused agents with this comprehensive, up-to-the-minute information unlocks immense potential for them to collaborate effectively, optimizing business outcomes and achieving strategic objectives.