The rise of autonomous AI agents presents a dual challenge: scaling agent throughput and safely running untrusted, LLM-generated code. This session will explore techniques for building agent infrastructure that address both performance and security while bringing Gemini’s power to AI code-execution use cases. We’ll demonstrate how to enable subsecond cold starts for massive concurrency, and how gVisor’s kernel-level isolation provides a zero-trust environment to protect your infrastructure. We’ll also examine how LangChain, a leading AI agent platform, is leveraging this architecture to deliver autonomous agent experiences at massive scale.