Your agent works in the demo. Then it runs for nine minutes, calls twelve tools, waits on a human, and crashes at step seven. The notebook architecture quietly falls apart. This talk is about why: agents are long-running stateful workflows, but we deploy them on short-lived stateless infrastructure. That mismatch kills more agents than bad prompts ever will.

Serverless runtimes assume work finishes fast and forgets between calls. Agents violate both. They block on slow tools. They wait on humans. They hit execution limits mid-thought. They get retried-and replay side effects they already committed. The failures are distributed-systems failures: lost progress, duplicated actions, orphaned runs.

I'll share the patterns we use to run millions of agent executions on serverless, applicable to Lambda, Cloud Run, Kubernetes, or Cloudflare Workers:

- Why agents break statelessness - wall-clock limits, cold starts, requests vs. workflows
- Durable execution - checkpointed steps that resume instead of restart
- Idempotency - making tool calls safe to replay
- Queue-orchestrated handoffs - decomposing "one big prompt" into recoverable units
- Human-in-the-loop as state - pausing for hours without holding a connection
- Per-tenant observability - debugging step 7 of 12 at 2am across thousands of runs

You'll leave with a vendor-neutral model for resumable agents, the failure modes to design against, and a live demo: we kill an agent mid-run and watch it resume from checkpoint.