Clare Liguori on AWS Strands SDK's Model-Driven Agent Design

In this SE Radio episode, AWS Senior Principal Engineer Clare Liguori discusses the design philosophy behind the open-source Strands Agents SDK. She frames any agent around three core pieces: a model, a set of tools, and a prompt. The SDK grew out of her work on Amazon Q Developer, where early unreliable models forced teams to build heavy scaffolding — custom parsing, feedback loops, and prompt pipelines — just to get consistent tool selection.

The turning point came as models like Claude Sonnet 3.5 got dramatically better at tool selection: the team realized that the scaffolding built up around earlier, weaker models — proactive RAG injection, mega prompt pipelines, manual response parsing — was now constraining rather than helping. Shifting to a "model-driven" approach, where RAG becomes just another tool the model chooses to invoke, measurably improved both accuracy and agent quality.

The conversation also covers MCP integration, sub-agent and multi-agent patterns, memory layers, production readiness, evaluation strategies starting from deterministically-scorable use cases, and common anti-patterns for engineers new to agent building. It's a practical look at when to trust the model versus when orchestration still adds value in production agent systems.