Agentic AI completes tasks on a user's behalf by executing functions, known as tools, to interact with an application's data and workflows.
An AI can either execute a task directly on the server for efficiency or perform the steps in the client-side UI to teach the user the workflow.
Use a Pinia store to manage a global tool registry and a composable to register or deregister tools based on a component's lifecycle.
Create well-designed tools by using specific names, returning useful results with clear error messages, writing precise descriptions, and providing UI state as context.
Implement a spectrum of trust by requiring explicit user confirmation for destructive or high-risk tool calls to prevent unintended consequences.
The system prompt acts as onboarding documentation for the AI, allowing you to encode UX decisions and behavioral rules in natural language.
Building agentic AI involves solving for challenges like tool explosion, latency, non-determinism, and developing new strategies for testing probabilistic components.
The future of agentic AI lies in moving beyond chat boxes to create smarter, context-aware actions that are deeply integrated into the application's UI.
Jobs that call for the skills explored in this talk.