Right now, most agents are a for loop with anxiety: call a model, wait, call a tool, wait, call another agent, pray. It's a beautiful demo and a production incident waiting to happen. One timeout, one rate limit, one flaky tool, and the whole synchronous chain collapses — taking your workflow (and your on-call weekend) with it.

Here's the uncomfortable truth: an agent isn't a script, it's a distributed system. And cloud native already solved distributed systems a decade ago. In this talk I'll rebuild an agent as an event-driven system on Kafka/NATS, where each step is an event, retries and idempotency are first-class, and long-running workflows survive failure instead of falling over. I'll walk through a working reference implementation, the failure modes that forced each decision, and the trade-offs you're signing up for — latency, complexity, debuggability.

You'll leave able to stop for-looping your agents and start engineering them.