Most AI agents are introduced as a loop: think → call a tool → observe → repeat.

That model works beautifully in a demo. Then the process crashes halfway through a task. A tool times out after performing the operation. Two workers pick up the same job. The model retries a non-idempotent action. A deployment happens while an agent is running. Or an hour-long task fails at minute fifty-seven and starts again from scratch.

At that point, the interesting problems are no longer LLM problems. They are distributed-systems problems.

This session looks at production agents through that lens. We will map familiar concepts—queues, leases, heartbeats, idempotency keys, checkpoints, durable execution, retries, cancellation and recovery—onto real agent architectures.

The goal is not to turn every agent into Kubernetes. It is to identify exactly where classical distributed-systems thinking becomes necessary once an agent moves from a synchronous demo to software expected to run reliably for minutes, hours or days.

By the end, the audience should be able to look at an agent architecture and immediately identify where it can duplicate work, lose progress, corrupt state or fail unrecoverably.

Target audience: Backend engineers, platform engineers, AI engineers, architects and anyone shipping agents beyond prototypes.

Preferred duration: 35–45 minutes. Can be adapted to a 20-minute condensed version.

Technical level: Intermediate. Familiarity with backend systems is useful; no ML background required.

Format: Architecture diagrams, failure scenarios and implementation patterns. Optional live demonstration of an interrupted agent recovering from a checkpoint.

Product/vendor neutrality: The ideas apply across agent frameworks and model providers.

First public delivery: Available as a first public delivery for the selected conference/community.