New in Confluent Intelligence: Real-Time Context Engine Upgrade, New Model Support, ML Functions, and More

As AI models become increasingly interchangeable, what matters isn’t which large language model (LLM) you choose – it’s whether your agents can see and act on the live state of your business. Context is the real competitive advantage. And when that context is stale, fragmented, or locked behind brittle point‑to‑point integrations, even the best models fail to deliver reliable decisions in production.

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

Today, we’re moving Confluent Intelligence into its next phase with a powerful Real-Time Context Engine generally available (GA), expanded agent operations, more model options, and new ML functions for real-time AI.

Q2’26: What’s New in Confluent Intelligence

We’re excited to announce the following new capabilities:

• Real-Time Context Engine (GA): Upgrade includes low‑latency enhanced querying capabilities such as filters, ranges, compound queries, projections, and ordering, so agents can access rich context without relying on external databases.

• Streaming Agents (GA): Production‑ready, event‑driven agents running natively on Flink and Kafka, now with GA support and enterprise‑grade operations.

• Agent Management Console (GA): Centralized, UI-driven control plane for creating and operating Streaming Agents at scale.

• Additional model support:

  • TimesFM (Early Access [EA]): Time-series forecasting models embedded into streaming pipelines.

  • Anthropic (GA): Native support for Anthropic models.

  • Fireworks AI (GA): Access to a broad catalog of optimized foundation models.

• Built-in ML functions:

  • Multivariate anomaly detection (OP): Moved from EA to Open Preview for detecting complex, multiple correlated metrics.

  • PII detection (EA): Automatically detect and redact sensitive fields in real time.

  • Sentiment analysis (EA): Score sentiment on streams of events and text for real-time customer service and operations.

Let’s dive into key areas in more detail.

Real-Time Context Engine (GA): Rich Queries Directly on Live Tables

Real-Time Context Engine continuously serves fresh context to any AI agent or application at low latency via MCP. Until now, it focused on primary key lookups: blazing fast, row‑by‑row retrieval for agents. With GA, it expands its role as a low‑latency context engine to let you query tables far more flexibly—without standing up and managing separate databases.

At GA, Real-Time Context Engine adds the following:

• Enhanced query support at low latency for filters, ranges, compound queries, projections, ordering, and more in motion. 

• Scale to any size: Designed to scale with the volume and cardinality of your streaming data, so growing traffic doesn’t force you into separate operational databases.

• Full schema support:

  • All schema types: AVRO, JSON, and Protobuf, with schema evolution support.

  • Nested values: First‑class support for nested and complex structures.

  • All data types: Coverage across character, string, integer, etc.

• Terraform and UI support for control plane operations:

  • Enable or disable topics feeding Real-Time Context Engine.

  • Update tool descriptions used by agents and other MCP clients.

  • Manage table exposure and life cycle programmatically as part of your platform automation.

For most teams, the hardest part of production AI is context engineering—assembling the right slice of fresh, governed data at decision time. Learn more from the ebook The Complete Guide to Context Engineering for AI and get started with docs here.

Use Case Example: Real-Time Credit Decisioning

A bank may have tables for customer profiles and risk segments, recent transaction windows, device fingerprints, and geo patterns. When a new credit application arrives, an agent can do the following:

1. Use Real-Time Context Engine for access to fresh context and advanced querying on the fly

2. Use the built-in ML function for anomaly detection to find application inconsistencies

3. Call an LLM directly using Flink SQL for explanation generation

4. Make an approve/decline/needs‑review decision in real time, not hours or days later

Streaming Agents (GA) and Agent Management Console (GA): Production-Ready, Event-Driven AI

Streaming Agents bring agentic AI directly into your data streams. Instead of polling data warehouses or relying on batch snapshots, event-driven Streaming Agents continuously monitor live business signals and take autonomous action in real time.

With GA for Streaming Agents, you now get enterprise‑grade reliability, with a four nines SLA, production support, and consistent runtime behavior for long‑lived agents consuming high‑throughput streams. There’s also an agent reflection pattern that lets agents iteratively critique and refine their own outputs before emitting a single, trusted event into the stream. Start building your own event-driven agents in minutes with the Quickstart and docs.

Now with Agent Management Console, you gain more operational control for Streaming Agents in the Confluent Cloud UI. The Agent Management Console pulls this into a single, visual experience. Streaming Agents show up as first-class resources, so developers and platform teams can see how an agent is wired—its inputs and outputs, prompts, models, tools, tables—without digging through code.

The new Agent Management Console enables teams to do the following:

• Make agents visible and manageable. See all agents, their statuses, and their core configurations in one place instead of going through SQL and jobs.

• Accelerate iteration. Create and refine prompts, models, tools, and data wiring through a guided console, not just code.

• Operate with confidence. Test agents with sample inputs, monitor live runs, and inspect logs to improve accuracy, latency, and reliability.

In practice, the Agent Management Console becomes a place where AI, data, and platform teams can collaborate on how agents are configured and how they behave in production, using the same rigor they already apply to other core services on Confluent Cloud.

Because everything is event‑driven and replayable, you can iterate on pipeline and agent logic with full observability and auditability.

Multivariate Anomaly Detection (OP), PII Detection (EA), and Sentiment Analysis (EA): ML Made Easy

In Q1'26, we introduced multivariate anomaly detection as an open preview built‑in ML function to find anomalies across multiple correlated metrics at once. This quarter, it reaches OP, and we’re adding two new ML functions in EA that are focused on trust and customer experience.

Multivariate Anomaly Detection

Multivariate anomaly detection lets you treat multiple signals as a single vector and detect when their combined behavior looks abnormal, not just when a single metric crosses a static threshold. 

Use cases include:

• Fraud detection: Combine transaction amount, merchant category, device, and location into a single feature vector to flag risky behavior.

• Predictive maintenance: Monitor temperature, vibration, and pressure across industrial equipment to catch early failure patterns.

• Customer service operations: Detect multi‑signal anomalies across page views, cart behavior, and latency that indicate real conversion problems.

PII Detection and Redaction

The new ML function for PII detection and redaction helps you protect sensitive data in motion. It detects common forms of PII (e.g., names, addresses, phone numbers, emails, IDs) in streaming data as it flows through Kafka, flags or redacts sensitive fields in real time, and then emits clean and compliant data into downstream topics. Instead of building custom solutions or moving data to warehouses, teams can now protect PII with a single SQL function call. 

Use cases include:

• AI safety: Redact PII before text or events are sent to AI pipelines and external LLMs.

• Data governance: Enforce consistent redaction policies across teams and applications.

• Regulatory compliance: Flag violations for compliance teams and ensure that sensitive values never leave boundaries in clear text.

Sentiment Analysis

The new sentiment analysis ML function runs on streaming data to score the tone of customer‑facing text in real time. It can optionally break down sentiment by specific aspects (e.g., cost, quality, wait time, support). By emitting structured sentiment signals directly into Kafka topics, it enables teams to detect what users are unhappy about as events flow through the system. This turns sentiment from batch jobs into something you can act on immediately. Learn more from docs here.

Use cases include:

• Live customer experience monitoring: Detect negative sentiment about wait time or support while users are still active.

• Incident and escalation workflows: Trigger alerts or routing based on sentiment changes in tickets or chats.

• Product and pricing feedback loops: Catch immediate reactions to launches, promotions, or outages.

• Agent-driven systems: Feed structured sentiment signals directly into Streaming Agents or use as part of an anomaly detection pipeline on Confluent Intelligence.

Sign up to try PII detection and sentiment analysis while these ML functions are in EA. 

TimesFM, Claude, and Fireworks AI Support

Confluent Intelligence is intentionally model‑agnostic: You bring your models, and we provide the real‑time, governed context and streaming infrastructure. In Q2, we’re adding new options across time‑series forecasting and LLMs.

TimesFM (EA): Time-Series Forecasting in Motion

With TimesFM support, you can call best-in-class time-series forecasting models directly from your streaming pipelines to forecast traffic, demand, or resource usage in real time as events arrive. 

The new AI_DETECT_ANOMALIES function is powered by TimesFM 2.5, a foundation model from Google Research that’s purpose-built for time-series forecasting. It performs in-context learning, allowing you to provide past data that steers the model toward a forecast and prediction. It differs from ML_DETECT_ANOMALIES in that the existing ML_DETECT_ANOMALIES uses ARIMA-based statistical forecasting, a solid and traditional approach. The new AI_DETECT_ANOMALIES steps it up by using a managed foundation model trained on massive time-series datasets, giving you more sophisticated pattern recognition out of the box. Think of it as going from a reliable calculator to a model that actually understands time-series behavior. Learn more from docs here.

You can use forecasts to trigger Streaming Agents to take autonomous action (e.g., autoscaling, replenishment, dynamic pricing). 

Anthropic Claude (GA) and Fireworks AI (GA): Flexibility with Security

We’ve expanded our Flink-based remote model support. Teams can now integrate their own Anthropic and Fireworks AI accounts directly into their stream processing workflows to build sophisticated, real-time AI applications.

• Anthropic support: Harness Claude’s advanced reasoning directly in your data streams. (View docs.)

• Fireworks AI Support: Access inference-optimized open source models (e.g., Llama 3, Mistral) with minimal latency. (View docs.)

The benefits include more streamlined and cost-efficient AI pipelines, since models can be called directly from Flink SQL using your credentials rather than through a cloud service provider (CSP). Model inference together with context from the Real-Time Context Engine, vector search, and built‑in ML make it easy to ensure that outputs are always grounded in fresh, trustworthy data. 

Confluent Intelligence provides model flexibility to ensure that teams use the right tool for the job. Whether you need the reasoning of Claude or the lightning-fast performance of Fireworks, you can use either or both as part of AI pipelines. Finally, models remain secure and sovereign; by using your own accounts, you maintain full control over your API keys, rate limits, and model configurations.

Start Building Real-Time, Context-Rich AI with Confluent Intelligence

To turn AI from a set of siloed experiments into production systems, you need Streaming Agents that act on operational events as they happen, not snapshots. Real-Time Context Engine continuously serves trustworthy context to your AI applications while AI/ML functions help detect complex anomalies, protect PII, and understand sentiment in real time. This is on top of a model-agnostic ecosystem with secure networking across your cloud provider of choice. 

Confluent Intelligence brings all of this together on a fully managed platform powered by Kafka and Flink, so you can focus on building AI that moves your business forward.

Get started today:

• Sign up for Confluent Cloud to enable Confluent Intelligence on your clusters and topics.

• Enroll in EA for TimesFM, PII detection, and sentiment analysis to apply ML directly on your data streams—no coding or ML expertise required.

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.