SE Radio 728: Clare Liguori on AWS Strands SDK for AI Agents

Building effective AI agents requires a shift from complex, brittle orchestration workflows toward a model-driven approach that leverages the inherent capabilities of modern large language models. By defining agents through the core components of models, tools, and prompts, developers can eliminate unnecessary scaffolding and allow models to handle tool selection and context retrieval dynamically. Implementing steering hooks provides a deterministic layer of validation, ensuring process adherence and reducing hallucinations by verifying tool inputs and outputs in real-time. Integrating the Model Context Protocol (MCP) further standardizes tool ecosystems, enabling seamless connectivity across internal APIs. As agent development matures, techniques like AI functions—which generate and execute code at runtime with defined pre- and post-conditions—offer a path toward more reliable, self-correcting systems that improve performance without requiring manual, step-by-step procedural logic.