Confluent Inc., maker of a platform for building and managing real-time data streaming pipelines using Apache Kafka, is offering a boost to organizations in financial services, healthcare, advertising and other time-dependent industries with today’s launch of Streaming Agents, a new capability in the company’s Confluent Cloud for Apache Flink.

The feature is intended to help organizations move beyond generative AI experiments and into scalable, production-grade agentic AI applications that take autonomous actions on real-time data.

Streaming Agents brings live data into data processing pipelines so organizations can build agents that reason, adapt and take immediate action as business conditions change.

“Agentic AI has become an important element of competitiveness, but what separates enterprises getting real ROI from [foundation models] is the ability to contextualize AI with domain-specific data in real time,” said Andrew Sellers, Confluent’s head of technology strategy.

Confluent cited International Data Corp. research that found enterprises ran an average of 23 generative AI proof-of-concept projects between 2023 and 2024, but only three reached production and fewer than two-thirds met expectations.

“[Large language models] aren’t the inhibiting force anymore,” he said. “What’s missing is contextual, trustworthy, real-time data that agents can use to make decisions.”

Continuous response

Streaming Agents integrates Flink-based stream processing with agentic frameworks, enabling event-driven agents to operate continuously, respond to signals and utilize tools such as databases and application programming interfaces to complete tasks. Confluent said such systems closely emulate how humans respond to dynamic environments.

For example, in financial services, anti-money-laundering workflows are often triggered by a series of suspicious transactions. “An agent can ingest those events, recognize patterns, generate a report using templates from past cases and hand that to an investigator,” Sellers said.

The pairing of Kafka and Flink is central to Confluent’s approach. Kafka provides durability and replay capabilities for messages passing between agents and systems, while Flink supports long-running workflows, watermarking, SQL and checkpointing for complex event processing.

“We think of this as analogous to the microservices era,” Sellers said. “Microservices only became compelling when they were event-driven and replayable. The same is true for AI agents.”

MCP support

Streaming Agents use the Model Context Protocol to securely invoke external tools like APIs and applications. “We are big fans of MCP and other technologies that make [messaging] more accessible,” he said. “MCP is quickly becoming as close to an industry standard as we have in this space.”

With Flink as a backbone, Streaming Agents support secure integration with vector databases, LLMs and other back-end systems. “We’re not trying to create competing standards or lock-in,” Sellers noted. “We’re augmenting the tools developers already have. They can keep doing what they’re doing and add streaming to enhance context and responsiveness.”

Agents can query and enrich real-time Kafka streams with data from relational databases or Representational State Transfer or REST APIs for better decision-making in applications like retrieval-augmented generation. Developers can also test agents using real data without causing unintended side effects, enabling dark launches, A/B testing and safer iterations, Confluent said.

All features are available now in open preview for Confluent Cloud users.

Image: SiliconANGLE/Microsoft Designer