Most “agent chains” today are just synchronous function calls with better branding. They look elegant in a notebook, but fall apart the moment you care about scale, cost, or correctness. Retries explode. State leaks into prompts. Failures become non-deterministic. And when something goes wrong, there is no way to replay what actually happened.

This talk takes a different approach. Instead of treating agents as functions, we treat them as event-driven, stateful processes, coordinated through streams. Apache Flink becomes the orchestration layer—not by adding a new framework, but by using the primitives it already provides: keyed state, timers, and exactly-once execution.

We’ll start by looking at why chained agents fail in production. The problem isn’t the models; it’s the control flow. Synchronous chains hide time, collapse state into text, and make cost an emergent property rather than an explicit one. Streaming forces those concerns into the open. Time becomes part of the model. State becomes durable. Retries become deliberate.

From there, we’ll walk through how to model agents as Flink operators. Each agent consumes events, maintains explicit memory, and emits decisions as new events. Chains form naturally through streams rather than calls, allowing fan-out, fan-in, partial completion, and human-in-the-loop intervention without breaking the system. Timers handle deadlines and escalation. State handles memory. Exactly-once semantics make the whole thing replayable.

The result is an agent system that behaves like production software, not a demo. You can enforce budgets. You can audit decisions. You can replay failures without re-running models. And you can reason about what the system is doing right now—not just what it did last night.

We’ll ground this in real examples: multi-step support agents, automated operations workflows, and decision pipelines where models, humans, and systems all act as peers. No hype, no new abstractions—just a practical way to build agent systems that are correct under pressure.

If you’re experimenting with agents and wondering how they survive contact with reality, this talk is about how streaming turns fragile chains into something durable.