Inside Flink Agents: Open Source Agents for the Enterprise

Flink Forward Barcelona 2025 was not just about streaming. It was about what comes next for enterprise AI. I sat down with Qingsheng Ren, Team Lead, Flink Connectors & Catalogs at Ververica, and Xintong Song, Staff Software Engineer at Alibaba Cloud, to talk about something that could change how enterprises build AI systems in production: Flink Agents.

Flink Agents is being introduced as an open source sub-project under Apache Flink. The goal is simple and ambitious at the same time: bring agentic AI into the same reliable, scalable, fault-tolerant world that already powers real-time data infrastructure.

We talked about why this matters.

First, why Flink Agents and why now?

They walked me through the motivation. Most AI agent frameworks today look exciting in a demo, but they break once you try to run them against live data, streaming events, strict SLAs, audit requirements, cost pressure, and real users. There’s a big gap between prototypes and reliable operations. That’s the gap Flink Agents is aiming to close.

Why open source?

Both Ververica and Alibaba made it clear that this is not meant to be a proprietary, closed feature. They want this to be a community effort under Apache Flink, not a vendor lock-in story. The belief is that enterprises will only bet on AI agents at scale if the runtime is open, portable, and battle tested.

How is building an AI agent different from building a normal Flink job?

This part was interesting. A standard Flink job processes streams. An agent has to do more. It has to reason, take actions, call tools, maintain context, react to feedback, and keep doing that continuously. You’re not just transforming data. You’re orchestrating behavior. Flink Agents is meant to give you those building blocks on top of Flink instead of forcing teams to stitch this together themselves.

What kind of companies is this for?

We got into enterprise workloads that actually need this. Think about environments where fast decisions matter and you can’t afford to go offline:

-- Fraud detection and response

-- Customer support and workflow automation

-- Operational monitoring, alert triage, and remediation

-- Real-time personalization and recommendations

-- Anywhere you need an autonomous loop, not just a dashboard

-- And finally, roadmap.

We talked about the next 2 to 3 years. The focus is on deeper runtime primitives for agent behavior, cleaner developer experience, and patterns that large enterprises can trust and repeat.

My takeaway:

Flink Agents is not just “yet another agent framework.” It’s an attempt to operationalize agentic AI on top of a streaming backbone that already runs at massive scale in production.

This is the conversation every enterprise AI team needs to be having right now.

#FlinkForward #Ververica #Streaming #RealTime #DataEngineering #AI #TheRavitShow