# New in Confluent Intelligence and AI Tools: Making Agents Native to the Stream, Expanded Model Support, New Agent Skills, and Copilot > Source: > Published: 2026-08-18 14:00:10+00:00 Learn why trusted, real-time data is critical for scaling AI and driving business value | [Learn More](https://ibm.webcasts.com/starthere.jsp?ei=1770442&tp_key=afb6aac3ff&sti=web) A customer writes in asking where their order is. The AI support agent checks the account, sees the order marked shipped, and sends a reply. The order was cancelled forty minutes ago. The agent wasn't wrong about anything it could see. It was reasoning over stale data that refreshes every six hours. This is why AI projects stall. Not because the models aren't capable, but because they lack AI-ready data and a reliable view of the current state of the business. Here is the promise of **Confluent Intelligence**: Build real-time, context-aware AI systems on Apache Kafka® and Apache Flink® that can monitor and take action on live business signals. By unifying streaming, data processing, context serving, inference, and AI / ML workflows in a governed way, Confluent Intelligence helps teams move AI beyond experiments into production – without stitching together disconnected systems. This quarter, we’re expanding that foundation in two important directions. First, Confluent Intelligence makes agents more native to streaming with anomaly detection and forecasting with expanded model support, including IBM Granite Time Series models and Google TimesFM model. Second, lightning queries and Real-Time Context Engine make it easier than ever for AI agents, applications, and analytics tools to work with fresh, structured views of what’s happening in the business right now. We’re also making Confluent Cloud easier to build on with expanded **AI tools**. Confluent’s fully managed MCP server lets AI understand and act on data and infrastructure in Confluent Cloud, with the ability to configure and restart fully managed connectors. New Agent Skills package Confluent expertise for AI coding assistants that developers already use, while Confluent Copilot provides a hosted, natural-language experience for building faster and operating Confluent Cloud. Together, these updates ground enterprise AI in always fresh business context and best practices, empowering agents to take action when it matters. Today, we’re announcing these new capabilities: **Real-Time Context Engine (GA):** Upgrade includes metrics enhancements. **Expanded model support:** **IBM Granite time-series models TTM, FlowState, PatchTST-FM (Early Access [EA]): **Robust time-series anomaly detection and forecasting models. **Google TimesFM model (EA):** Time-series forecasting model. **Fully Managed MCP Server updates (GA): **Write tools for connector restart and update connector configuration available now, and upcoming Flink tools support in later Sept. **New Agent Skills (GA, open source):** Migrate, Java Kafka producer/consumer, Flink UDFs, Confluent Skill creator. **Confluent Copilot for Confluent Cloud (EA):** Pre-built, Confluent-hosted copilot helps teams work faster, without configuring MCP servers or depending on a third-party AI assistant. Let’s take a closer look at each feature. AI is only as useful as the context it can access at decision time. [Real-Time Context Engine](https://confluent.io/product/real-time-context-engine) continuously turns streaming data into structured context that agents and applications query through MCP, so they reason over what is happening now instead of a stale snapshot from the last warehouse refresh. Because the context is served directly from Confluent, where the data is already moving, there is no separate operational database to keep in sync and no point-to-point integration to maintain. The context is durable and independent of the topics it derives from, so a stream can age out while the context it produces stays queryable for as long as agents need it. [Context engineering](/resources/ebook/complete-guide-to-context-engineering-for-ai/) becomes a managed capability with governance already attached, rather than plumbing that every team builds and owns. With enhancements in Q3, Real-Time Context Engine now surfaces user-facing metrics around usage, performance, health, and cost visibility, including: **Table size:** how much data is being maintained **Scan rate:** how much data AI agents and applications scan when retrieving context **Throughput and query rate:** sustained write throughput and queries per second This makes AI context operable and predictable in production where metrics API and billing updates provide clear visibility into usage and cost drivers, helping you monitor production workloads, forecast spend, and scale AI with greater confidence. Learn more from [docs here](https://docs.confluent.io/cloud/current/ai/real-time-context-engine.html) and get started with Real-Time Context Engine today. Real-Time Context Engine exposes fresh business context for AI through MCP. Beyond MCP, we’re announcing lightning queries, which give developers and analytics teams a REST API path to execute low-latency queries against the same live state. Together, they extend the value of Confluent’s data streaming platform from capturing and processing data to serving it where decisions are made – without replicating it into a separate operational database or managing more data pipelines and data syncs. That shared serving layer supports a range of production use cases. For example, an AI customer-service agent can monitor delayed orders and determine the right recovery action, such as offering a credit or expediting a shipment, using the latest order and fulfillment state. A merchandiser can query current inventory risk across stores to identify products likely to stock out in the next two days. And a customer-facing mobile application can check whether a product is available for pickup within 10 miles of a customer’s zip code – all with fast access to data that is already flowing through Kafka. Lightning query is now available in Early Access.[ Sign up to try it](https://events.confluent.io/early-access-flink-features) and experience the future of real-time data serving for AI, analytics, and operational applications. Confluent is model-agnostic: You select the models that best fit your use case, and Confluent provides the fresh data and streaming infrastructure to put them to work. In Q3, we’re expanding model flexibility and choice with new options across time-series models: IBM Granite Time Series and Google TimesFM for anomaly detection and forecasting. Anomaly detection and forecasting are most valuable when they run close to the moment a business changes – not after data has been copied into a warehouse or downstream system. Teams can call IBM Granite time-series models and Google TimesFM directly in Flink to detect unusual behavior and forecast what is likely to happen next on streaming data. This makes it easier to turn continuously arriving events into actionable insight without building and operating a separate ML stack. **IBM Granite Time-Series Models (TTM, FlowState, PatchTST-FM)** IBM Granite time-series models are used for both forecasting and anomaly detection. All three Granite models – TTM, Flowstate, and PatchTSTFM – are selectable in Confluent’s existing `AI_FORECAST` and `AI_DETECT_ANOMALIES` Flink SQL functions. The integration allows you to choose between models with a single SQL parameter: same interface, with more model choice and no pipeline redesign. Different foundation models can perform better on different data profiles. For example, TimesFM may be a strong fit for some workloads, while Granite models may be better suited for manufacturing sensor data, industrial IoT, or financial time-series data. Model choice lets teams benchmark alternatives and select the best fit for their workload, while IBM Granite also brings IBM’s enterprise AI governance framework for organizations that need model provenance and licensing transparency. **Google TimesFM Model** TimesFM support brings a best-in-class managed forecasting model directly into Flink. With `AI_FORECAST` , you can predict future values from streaming time-series data with confidence intervals. With `AI_DETECT_ANOMALIES` , you can flag values that fall outside expected bounds as events arrive. These forecasts can drive Streaming Agents to take action – such as replenishing inventory or adjusting prices – without requiring a separate model-serving stack. Expanded model support gives teams more choice without adding infrastructure. The models have **zero configuration **– hosted in Confluent Cloud and called directly from Flink, there is no setup and credential management while keeping data in Confluent Cloud ensures security. Shared GPU resources and no cloud ingress or egress simplify the architecture and help control the cost of real-time AI workloads. For example, a team can forecast demand from live order events to prepare inventory before a stockout, or analyze trucking signals to identify a developing issue before it affects customers. IBM Granite models and Google TimesFM are available in Early Access on AWS. Learn more from [forecasting](https://docs.confluent.io/cloud/current/ai/builtin-functions/forecast.html) and[ anomaly detection](https://docs.confluent.io/cloud/current/ai/builtin-functions/detect-anomalies.html) docs and try it today using the [AI/ML Quickstart](https://github.com/confluentinc/quickstart-ai-ml-functions). A fully managed MCP server is a Confluent-hosted MCP endpoint, so teams do not need to deploy or maintain MCP infrastructure. Confluent’s fully managed MCP server give AI assistants permissioned access to Confluent Cloud without requiring teams to run a local MCP server. New MCP capabilities this quarter include write tools to restart connectors and update connector configuration, and Flink tools for creating and managing Flink SQL. Together, these MCP server updates give AI assistants a more direct way to work with Confluent Cloud resources and streaming applications. Get started with MCP in minutes using [docs here](https://docs.confluent.io/cloud/current/ai/ai-tools/managed-mcp-server.html). AI coding assistants can generate code, but they do not understand the configuration details, security practices, and operational patterns that make a streaming application production-ready. Confluent Skills package that expertise into reusable workflows for coding assistants (e.g., Claude Code, Cursor). They help developers turn natural language into a working Confluent solution with best practices, guardrails, and validation built in – reducing context switching and shortening the path from prompt to production. In Q3, new skills expand what developers can accomplish with Confluent, including: **Migrate:** The[ Migration skill](https://github.com/confluentinc/agent-skills/tree/main/skills/msk-migration) helps teams assess an MSK environment and plan a migration to Confluent Cloud. It guides users through environment discovery, risk identification, cluster sizing, and switchover planning for informed decision-making before execution. **Build Java Kafka clients:** The[ Java Kafka Client skill](https://github.com/confluentinc/agent-skills/tree/main/skills/developing-kafka-java-client) scaffolds production-ready producer and consumer applications with the Apache Kafka Java clients and Confluent Schema Registry serializers. It supports Maven or Gradle, Avro, JSON Schema, or Protobuf, and deployment to Confluent Cloud, local Docker, or WarpStream. **Extend Flink with custom logic:** The[ Flink UDF skill](https://github.com/confluentinc/agent-skills/tree/main/skills/flink-udf) helps developers build and deploy Java user-defined functions for Flink SQL and Table API – including scalar UDFs for value transformations, user-defined table functions for one-to-many operations, and process table functions for advanced stateful stream processing. **Create new Confluent skills:** The[ Confluent Skill Creator](https://github.com/confluentinc/agent-skills/tree/main/skills/confluent-skill-creator) is a skill for building skills. It guides contributors through defining scope, capturing user intent, following the Agent Skills specification, testing against a real Confluent environment, iterating with evaluations, and packaging the result. Together, these skills make AI coding assistants more useful for real streaming work: they bring Confluent knowledge into the developer workflow so teams can build with less guesswork.[ Install Confluent Agent Skills](https://github.com/confluentinc/agent-skills) with `npx skills add confluentinc/agent-skills` and learn more from [docs here](https://docs.confluent.io/cloud/current/ai/ai-tools/agent-skills.html). Confluent Copilot brings unique data streaming platform expertise into Confluent Cloud as a pre-built, Confluent-hosted experience: Describe what you want to accomplish in natural language, and Confluent Copilot will help you build, troubleshoot, and operate your streaming workloads without configuring an MCP server or relying on a third-party AI assistant. Grounded in Confluent documentation, APIs, troubleshooting knowledge, and best practices, Copilot is designed to accelerate onboarding and simplify operations while preserving user-scoped permissions, guardrails, and confirmation before actions are taken. A key use case focuses on fully managed connectors. When a connector fails or its performance degrades, Confluent Copilot can inspect connector status, errors, logs, configuration, and lag to identify likely root causes and recommend a fix. For example, it can help diagnose a failed Debezium source connector after a database maintenance event, or investigate rising lag caused by uneven task or partition load. Where appropriate, it can restart the connector or update its configuration – with explicit user confirmation before any change is executed. In addition, if an issue is due to something outside the connector (e.g., external system, networking), the copilot can provide guidance for remediation. This turns a workflow that can otherwise require days of documentation searches and support exchanges into a guided, Confluent-specific troubleshooting experience. [Sign up for Early Access](https://events.confluent.io/early-access-flink-features) to try Confluent Copilot today. AI delivers business value when it can understand what is happening now – and act on that context. Confluent Intelligence brings data streaming, processing, and real-time context serving into a single platform. With fresh context reflecting the current state of the business, flexible model choice, and capabilities for detecting anomalies and forecasting what comes next, you can build smarter AI that helps the business run better. AI Tools for Confluent extend that foundation by bringing Confluent expertise into AI coding assistants and providing a Confluent Copilot for building on, troubleshooting, and operating Confluent Cloud in natural language. Together, Confluent Intelligence and AI tools help organizations move from disconnected AI experiments to production systems that are grounded in current business reality and ready to make decisions in real time. Get started today: [Sign up for Confluent Cloud](/get-started/) to start building real-time AI on streaming data *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.* Model Context Protocol (MCP), introduced by Anthropic, is a new standard that simplifies AI integrations by providing a secure and consistent way to connect AI agents with external tools and data sources… Accelerate enterprise streaming and AI workloads with new product updates across Kora, connectors, Flink, Tableflow, AI anomaly detection and forecasting, security enhancements, and WarpStream.