Giselle van Dongen

In this session, we skip past the demos that work only on your laptop, and go straight to how you can build production-ready agents with a stack that covers all the hard bits of backend development that you don’t want to be bothered with when developing your agents: - Failure resiliency: retries, timeouts, and exactly-once execution so a flaky API or a crashed process doesn't corrupt your agent's state or makes them start from scratch - Durable Sessions: a session store with built-in conversation isolation and protection against corruption from concurrent agents - Pause/resume for human approvals: survive human approvals and research that take weeks without building complex infra - Agent-to-agent messaging layer: call agents developed by other teams or running on other infra with resilient HTTP calls - A kill switch: cancel a running agent cleanly at any point, without leaving half-executed work behind We will demonstrate each concept with live code examples, using Python, OpenAI Agents SDK and Restate as open-source Durable Execution engine. All examples are generally applicable: pick your favorite agent SDK (OpenAI Agents, Pydantic AI, Vercel AI, Google ADK,…) or go wild and implement low-level custom agents by just tying together LLM calls with custom logic.