New in Confluent Intelligence: A2A, Multivariate Anomaly Detection, Vector Search for Cosmos DB, Amazon S3 Vectors, and More

As AI models are increasingly commoditized, the value driver for enterprises is no longer “Which large language model (LLM) are we using?” but “How can we use our data for reliable, real-time AI decisioning?” Agentic AI systems—where agents plan, decide, and act autonomously—are only as useful as the context they have. When that context is stale, fragmented, or locked away behind brittle point-to-point integrations, even the best models fail to deliver.

Confluent Intelligence was created to solve this problem. It’s a fully managed service on Confluent Cloud for building real-time, context-rich, and trustworthy AI systems on a unified data streaming platform that brings together Apache Kafka® and Apache Flink®. With Confluent Intelligence, you can stream operational events, continuously enrich them with external data, apply built-in ML, and power Streaming Agents and other AI applications with fresh context—without stitching together infrastructure.

Today, we’re expanding Confluent Intelligence with new capabilities across Streaming Agents, built-in ML functions, and Model Context Protocol (MCP) support. Together, they make it easier to reuse the agents you already have, detect complex anomalies in motion, bring more vector stores into your retrieval-augmented generation (RAG) workflows, secure networking, and standardize how agents tap into real-time data on Confluent Cloud.

Q1’26: What’s New in Confluent Intelligence

We’re excited to announce new features and support for Confluent Intelligence:

• Agent2Agent (A2A) Integration for Streaming Agents (Open Preview): Enable Streaming Agents to collaborate and orchestrate tasks with external agents on any A2A‑capable platform (e.g., LangChain, CrewAI, SAP, Salesforce), all over a reliable, replayable Kafka backbone.

• Multivariate Anomaly Detection for Built-in ML Functions (Early Access): Detect anomalies across multiple correlated metrics at once, to flag real problems instead of noisy spikes.

• Vector Search for Azure Cosmos DB and Amazon S3 Vectors (General Availability): Bring Azure Cosmos DB and Amazon S3 Vectors into the same External tables and Search fabric you already use for MongoDB, Pinecone, Elastic, PostgreSQL, SQL Server, Oracle, and more, expanding real-time RAG pipelines in Flink.

• Amazon Web Services (AWS) and Azure Private Link for Model Inference, External Tables, and Search (General Availability): Securely call models and enrich real-time streams with sensitive data via private, VPC‑to‑VPC connectivity from Flink to external databases, vector stores, and REST APIs—keeping AI traffic off the public internet.

• Confluent Support for Open Source MCP Server for Confluent Cloud: Get vendor-backed support for the open source MCP server that provides governed, real-time data to any MCP client.

Let’s look at each of these in more detail.

A2A Integration: Reuse and Connect Your Existing Agents

As enterprises adopt AI agents across customer customer relationship management (CRM) systems, data warehouses, operational systems, and bespoke apps, they often end up with islands of agents—powerful in isolation, but unable to easily discover, communicate with, or coordinate each other’s work. Most of these interactions today are wired via point-to-point HTTP, which is hard to observe, audit, and scale as multi-agent systems grow.

A2A Integration for Streaming Agents brings the A2A open protocol directly into Flink, letting event-driven agents connect, orchestrate, and collaborate with agents on any A2A‑capable platform (e.g., LangChain, SAP, Salesforce) over replayable event streams on Kafka. You can keep the agents and frameworks you’ve already invested in; now they can become event-driven and act on the live state of your business instead of stale batch snapshots. Learn more from docs here.

Streaming A2A Use Case: Orchestrating an Event-Driven Incident Workflow

Imagine you’re running telecommunications incident management across Confluent, ServiceNow, Salesforce, and a custom remediation agent:

1. A Streaming Agent monitors telemetry Kafka topics for high-priority alerts (e.g., spikes in dropped calls in a specific region).

2. When a network anomaly is flagged, the Streaming Agent uses A2A to:

   1. Trigger a ServiceNow incident triage agent to enrich the ticket with recent logs, configuration changes, and impacted 4G/5G cell sites.

   2. Call a Salesforce agent to pull in the latest account and service level agreement (SLA) context for affected enterprise and consumer subscribers. 

3. As each external agent completes its work, results are written back to Kafka and streamed into a remediation agent, which decides whether to reroute traffic, send targeted customer communications, or escalate to a human.

Because this real-time, event-driven workflow is wired through Confluent Intelligence and A2A, you can replay the entire multi-agent interaction as part of an immutable log, inspect every decision, and iterate safely on new logic without impacting production.

Multivariate Anomaly Detection: Catch What Simple Checks Miss

Real-world systems rarely fail due to a single metric crossing a threshold. A normal card transaction may be suspicious only when paired with a new device, unrecognized location, or unusual spending velocity across multiple accounts. Traditional single-variable anomaly detection often misses these correlations—or distracts teams with false positives when noise or outliers skew the baseline.

Multivariate Anomaly Detection for Built-in ML Functions changes this by letting you treat multiple metrics as a single vector and applying robust statistics to detect when their joint behavior looks abnormal. This unlocks new types of real-time AI use cases by finding anomalies across different combinations of data, which is a far more powerful way to address complex scenarios that require more than just one signal. The result is minimized noisy alerts and earlier detection of real issues that matter to customers and the business. 

This feature uses the new ML_DETECT_ANOMALIES_ROBUST function in Flink SQL. Sign up for Early Access here and learn more from docs here.

Multivariate Use Case: Predictive Maintenance on Industrial Equipment

Consider predictive maintenance on a fleet of pumps. Failures may not be explained by a single reading; temperature, pressure, and vibration need to be monitored together to accurately detect early signs of trouble.

With Multivariate Anomaly Detection in Confluent Intelligence, you can:

1. Stream sensor data from different plants into Kafka topics. 

2. Use Flink SQL with ML_DETECT_ANOMALIES_ROBUST on vectors like ROW(temp, pressure, vibration) to model normal operating behavior over time.

3. Flag events where the combination of metrics deviates from normal—even if each metric individually still looks within bounds.

4. Feed those anomaly events to trigger an event-driven Streaming Agent that enriches them with maintenance history and automatically opens or prioritizes work orders in your asset management system.

The same pattern applies to fraud detection, where you might model transaction amount, merchant category, device fingerprint, and location as a vector; or to ecommerce scenarios where page views, cart abandonment, and checkout latency together signal a real conversion issue versus a transient blip.

Vector Search for Cosmos DB and Amazon S3 Vectors: Enrich Your RAG Stack

Businesses today have data spread across operational databases, analytics systems, and multiple vector stores. Operational events might be streamed to Kafka, while embeddings live in Cosmos DB or S3 Vectors, forcing teams to stitch together ingestion, embedding, vector search, and inference with fragile point-to-point connections.

Flink vector search on Azure Cosmos DB and Amazon S3 Vectors closes this gap by making both systems first-class vector providers in Confluent’s external tables and search fabric—alongside MongoDB, Elastic, Couchbase, and Pinecone.

With this release, you can:

• Query Cosmos DB or S3 Vectors directly from Flink SQL to retrieve k‑nearest‑neighbor (kNN) results in real time, before prompting your LLM.

• Run a single streaming pipeline that handles ingestion, transformation, embedding creation, vector search, and model inference instead of chaining multiple microservices.

• Ground Streaming Agents and RAG applications in relevant context, mitigating hallucinations and improving answer quality.

• Avoid duplicating data into extra operational stores by querying vectors in place via external tables and search.

• Stay future-proof by using the same Flink and Streaming Agents patterns across all supported vector databases.

For Amazon S3 Vectors, you also get low-cost, durable vector storage that easily integrates with your S3 data lake and AWS services such as Amazon Bedrock—enabling Streaming Agents to combine semantic context retrieved from S3 with Bedrock models in the same Flink job. 

Learn more from docs here.

Vector Search Use Case: Real-Time Customer Support With Fresh Context

Suppose you’re building a customer support agent that needs the most relevant context at prompt time from knowledge base articles and recent call transcripts in Cosmos DB or S3 Vectors.

With Confluent Intelligence:

• A streaming agent monitors Kafka topics for incoming support tickets and user events.

• For each ticket, Flink SQL performs vector search against Cosmos DB or S3 Vectors to retrieve the most semantically relevant content.

• The streaming agent combines these vector search results with real-time streams (e.g., loyalty status, recent browsing and cart activity, order/refund history) and calls the LLM to create a hyperpersonalized recommendation.

Because all of this is running in the same Flink stream processing pipeline, you can unify AI and data processing workflows while avoiding the additional cost and sprawl of separate ingestion jobs, bespoke RAG services, and ad hoc synchronization between vector databases and data sources.

AWS and Azure Private Link for Model Inference, External Tables, and Vector Search: Secure Networking for AI

For many teams, the last mile of connecting AI workflows to sensitive systems of record isn’t about APIs or schemas. It’s networking. 

AWS and Azure Private Link for model inference, external tables, and vector search enables private, VPC‑to‑VPC connectivity from Flink to your external models and systems—so Confluent Intelligence (including Streaming Agents) can call LLMs and enrich real-time streams with sensitive, proprietary enterprise data from external databases, vector stores, and REST endpoints without going over the public internet. For many organizations, this helps unblock production deployments and ensures information security compliance.

Private Link enables you to:

• Keep AI traffic private and compliant: Use Private Link so that lookups against CRM, enterprise resource planning (ERP), vector stores, and REST APIs happen over private network paths.

• Safely join Kafka streams with systems of record: Give Streaming Agents fresh, complete context (e.g., customer records, orders, policies) while meeting strict security requirements.

• Reduce networking complexity: Replace ad hoc proxies and tunnels with standardized, cloud-native private connectivity backed by Confluent’s (Teams adopting Anthropic’s MCP want a simple, standardized way for AI agents to tap into Confluent Cloud—Kafka, Flink, connectors, Tableflow, Schema Registry, and more—using the open source Confluent MCP server. Until now, they’ve largely been on their own: deploying and operating the MCP server themselves, wiring it to each MCP client (e.g., Claude Desktop, Goose, Gemini CLI), and handling issues without formal vendor backing.

Learn more from the docs here.

Support for Open Source MCP Server for Confluent Cloud: Production Backing for Any MCP Client

Teams adopting Anthropic’s MCP want a simple, standardized way for AI agents to tap into Confluent Cloud—Kafka, Flink, connectors, Tableflow, Schema Registry, and more—using the open source Confluent MCP server. Until now, they’ve largely been on their own: deploying and operating the MCP server themselves, wiring it to each MCP client (e.g., Claude Desktop, Goose, Gemini CLI), and handling issues without formal vendor backing.

Confluent now provides official support for the open source MCP server for Confluent Cloud, turning it into a production‑ready, vendor‑backed bridge between your MCP-based agents and real-time, governed data on Confluent.

This support means you can:

• Give MCP agents direct access to fresh Confluent Cloud data and operations far beyond just managing Kafka topics—covering connectors, Flink, Tableflow, support for billing and metrics APIs, and more. This can also be used to debug issues like backpressure.

• Manage Confluent via natural language—configuring topics, running Flink SQL, and interacting with data infrastructure from MCP clients without code.

• Raise issues via GitHub or your account team, with Confluent engineering working to resolve them under defined SLAs.

Start Building Real-Time, Context-Rich With Confluent Intelligence

To capitalize on the full potential of AI, teams need AI agents and applications that can see, understand, and act on the live state of the business. Confluent Intelligence brings together streaming and stream processing for context engineering, ML, vector search for RAG, Streaming Agents, and open protocols such as MCP and A2A to make that possible on a fully managed, governed platform.

Get started with Confluent Intelligence to turn your AI initiatives into real-time production systems today!

Apache®, Apache Kafka®, Apache Flink®, Flink®, and the respective logos are 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.