AI agents are only as reliable as the models behind them. Most teams start by wiring an agent to a single LLM and calling it done. Then reality hits: rate limits, outages, cost spikes, and tasks where one model underperforms another. The teams building resilient agents in production aren't betting on one model. They're building across many.
This talk covers how to architect AI agents that route intelligently across open and closed LLMs. I'll walk through practical patterns for model selection at inference time: when to use a large frontier model versus a fine-tuned open-weight model, how to build fallback chains that maintain agent quality during provider outages, and how to use routing logic to optimize for cost, latency, and task-specific accuracy.
Using PyTorch ecosystem tools like vLLM for self-hosted open models alongside closed API providers, I'll show how teams are deploying agent systems that aren't locked into any single vendor or architecture. We'll look at real tradeoffs between dense and MoE open models for different agent subtasks, and why the most resilient agent architectures treat model selection as a runtime decision, not a design-time one.