Picture a rebooking agent at SFO on a foggy morning. Flight operations, seat maps and departures live in three systems that update at different times, so the agent seats a passenger in a seat someone else already has. Shaun Clowes, who now runs Confluent and data at IBM, opened Current 2026 with that story. He called it “a different version of now.”
Our reading of the conference, held November 4-5 in San Francisco: event-driven architecture is how you give AI one consistent version of now, and Confluent shipped a real answer for the read side. Agents can query live business state straight from the streaming platform. Three problems remain yours. Agents that write to core systems still need their own safety layer, the new query feature runs only on AWS for now, and IBM Event Streams customers face a migration deadline.
In our Current 2026 preview we set six tests for the announcements. Below is how they held up.
Why event-driven architecture moved up the AI agenda
AI agents act on whatever data they can see. If inventory refreshes overnight and the agent promises same-day pickup at 2 p.m., the error reaches a customer before anyone looks at a dashboard. Batch pipelines were built for people who check reports a few times a day and make allowance for the lag. An agent makes no such allowance.
The numbers presented around the event point the same way. An IDC survey of 635 IT and data leaders, cited by CRN, found:
| Finding | Share |
|---|---|
| Organizations with real-time data movement in all workflows that need it | About 16% |
| Running multiple agentic apps in production, with limited real-time integration | 8% |
| Running multiple agentic apps in production, with comprehensive real-time coverage | More than 50% |
The survey shows correlation, and companies with strong real-time data probably differ in other ways too. Still, a gap of 8% versus more than 50% deserves attention. IBM CEO Arvind Krishna added that 90% of enterprise data is still outside the cloud, which explains why IBM bought Confluent: most of the data agents need lives in mainframes, ERPs and private data centers.
Accenture gave the most concrete example. About 80% of its data now flows over Kafka topics, and in one staffing application, requests that used to take days to reach the right people arrive within seconds. Notice the order of work. Accenture built the streaming backbone first and put AI on top of it later.
The enterprise problem behind Streamhouse
Companies trying to give agents fresh data usually end up with a patchwork of Kafka topics, Flink jobs, a connector and a Redis or Postgres cache, because Kafka itself cannot answer a question. Every hop adds delay, cost and another place for data to drift.
What Confluent shipped
Confluent’s answer is the Streamhouse, a data architecture that keeps the current state of the business available to applications, analytics and agents. Confluent says it works alongside existing warehouses and lakehouses rather than replacing them. To support it, Confluent added Query as the fifth pillar of its platform, next to Stream, Connect, Process and Govern.
| Query interface | What it does | Status |
|---|---|---|
| Lightning Tables | Low-latency SQL and API reads of current state, about 100 ms from event to query in Confluent’s testing | GA on AWS |
| Real-Time Context Engine | Serves that state to AI agents through MCP | GA |
| Tableflow | Streams into Iceberg and Delta tables for analytics | GA, now also on Google Cloud |
The practical gain is fewer moving parts: on AWS, an agent can look up an order’s live status without a separate cache to keep in sync.
Where the Streamhouse idea is still young
Five vendors formed a Streamhouse working group in September, and Current turned the term into a product. The lakehouse eventually settled on shared formats such as Iceberg. The Streamhouse definition so far describes goals, with no shared format behind it. Buy the capabilities if they fit, and assume that moving between vendors will take work.
Architecture and operating-model implications
A CFO who sees the first model bill from an agent wired to a busy event stream will ask what went wrong. The practitioner talks at Current gave a consistent answer: agents were placed in the wrong spot.
Keep agents off the hot path
Calling a language model on every event cannot work at enterprise volumes. Conduktor’s conference recap did the math: at 10,000 events per second, that is 864 million model calls a day. The pattern that works has three layers. Flink or Kafka Streams handle every event, validating, deduplicating, joining and computing. Small forecasting or anomaly models score the filtered stream and flag what looks unusual. Only then, for a delayed high-value shipment or a disputed charge, does an agent wake up, pull context and propose an action.
Most events never reach a model, and where a fixed rule covers the case, a workflow is cheaper and easier to audit than an agent.
Reading safely is not acting safely
Streamhouse fixes stale reads. Duplicate actions are a separate problem, and the session on them was the most useful one we watched. In Kafka as the Agent Action Log, Conduktor walked through a support agent that decides to refund $40 for a damaged blender, emails the customer and crashes before Kafka records progress. Kafka redelivers the ticket. The model is asked again, picks a replacement this time, and the company pays twice.
Kafka’s exactly-once guarantee covers writes between Kafka topics, not a call to a refund API. The fix is simple to describe:
- Write the agent’s decision to a Kafka topic before acting on it.
- Build each step ID from the source event ID, so a retry produces the same ID.
- On retry, replay the recorded decision instead of asking the model again.
- Pass the step ID to the target system as its idempotency key.
This is the events-versus-commands line from our preview. OrderCancelled records a fact. IssueRefund asks for an action, and a policy check should approve it before anything touches SAP or a payment system.
Operations will cost you more than brokers
Two operational points came up repeatedly. Conduktor’s Stéphane Derosiaux traced most Kafka support tickets to client-side habits, such as creating a new producer per request or tight retry loops, rather than broker bugs, so shared clusters need quotas before more teams join. Failover also depends on people: someone has to know which applications use which topics and who signs off on the switch.
Governance, integration and cost trade-offs
Pilots rarely fail on the demo. They fail in month three, when someone finds a limit that was in the documentation all along.
Read the documentation before the keynote slides
| Limit | What the docs say | Why it matters |
|---|---|---|
| Cloud | Lightning Tables run only on AWS, in selected regions | Azure-first or Google Cloud estates keep an external cache for now |
| SQL | Read-only queries on a single table, no joins or aggregates | Every join and metric must be built upstream in Flink |
| Schema changes | An unsupported change stops ingestion while the table still shows Active | An agent can read frozen data unless you monitor freshness directly |
Show the last row to your risk team. A status light that stays green while data stops updating is how an agent ends up acting on yesterday’s state. Confluent’s new Schema Registry compatibility policy rejects changes that would break Lightning Tables or Flink jobs, which helps if teams switch it on.
IBM Event Streams has a deadline
For IBM shops this is the most practical news of the season. IBM’s transition guide sets the dates for IBM Event Automation, which includes Event Streams and Event Processing:
| Milestone | Date |
|---|---|
| End of new orders | October 9, 2026 |
| End of renewal ordering | November 10, 2026 |
| End of normal support | April 30, 2028 |
| End of extended support | April 30, 2029 |
IBM points Event Streams users to IBM Confluent Platform and Event Processing users to Confluent Platform for Apache Flink. Event Endpoint Management stays and works with any Kafka-compatible platform. If you run IBM Cloud Pak for Integration, plan the inventory and migration waves in 2027. Many estates will keep IBM MQ next to Kafka, a trade-off we cover in our IBM MQ vs Kafka guide.
The bill has many meters
Confluent Cloud bills clusters, connectors, Flink, Tableflow and Lightning Tables separately, the last by compute time and data scanned. An agent that scans broadly on every question can make a cheap pilot expensive, so set query budgets first.
A decision framework for enterprise leaders
We sorted what was announced at or just before Current 2026 against our preview tests: business value, freshness, identity, action safety, recovery and cost.
| Verdict | Capabilities | Condition |
|---|---|---|
| Deployable now | Schema Registry compatibility rules, Tableflow for analytics, managed MCP servers with OAuth | Use OAuth instead of shared API keys; keep Tableflow for minutes-level freshness |
| Pilot with controls | Lightning Tables and Real-Time Context Engine on AWS; Event Streams-to-Confluent migration planning | Pre-join in Flink, monitor freshness, cap agent queries; start inventory now |
| Watch | Lightning Tables on Azure, Google Cloud or on-prem; on-prem Granite time-series models (open preview) | No date or GA yet |
| Build yourself | Safe write-back to systems of record | Action log, policy checks and idempotent APIs |
Vendor slides tend to leave out the last row. No Streamhouse feature makes an agent’s write to SAP or a payment system safe on its own.
Key takeaways and checklist
- Agents in production need real-time data. Batch feeds are now a risk for AI use cases.
- Streamhouse solves stale reads on AWS. Duplicate actions still need an action log.
- Put deterministic processing in front of every agent, and send only exceptions to a model.
- Monitor data freshness on the data itself, because table status can stay green.
- IBM Event Streams users have until April 2028. Start planning in 2027.
Before you connect agents to event streams, check:
- Each agent use case has a written freshness need, measured end to end.
- No language model sits on a high-volume topic.
- Joins and metrics are built in Flink before they reach a query table.
- Decisions are logged before actions, with deterministic step IDs.
- Agents use scoped OAuth identities, not shared API keys.
- Freshness alerts check the data’s last update time.
- IBM Event Streams and Event Processing deployments are inventoried.
- Query and compute budgets are set per agent.
Planning event streams for AI agents?
In a 30-minute call we look at your Kafka, MQ and integration estate and tell you which agent use cases are ready, and what to fix first.
One call, no commitment.
One call, no commitment.
FAQ
What is event-driven architecture?
Event-driven architecture is a design in which systems communicate by publishing and reacting to events, such as an order placed or a payment cleared, rather than calling each other directly. Producers and consumers stay decoupled, so each can change or scale on its own.
What is the difference between microservices and event-driven architecture?
Microservices describe how you split an application into small services. Event-driven architecture describes how those services, or any systems, communicate. Many microservice estates use events, but you can also apply event-driven patterns to ERPs, mainframes and SaaS tools.
Is Kafka an event-driven architecture?
No. Apache Kafka is a platform for storing and streaming events. It is often the backbone of an event-driven architecture, but the architecture also includes event design, schemas, processing, governance and the services that react to events.