Player not loading? Watch on YouTube
This walkthrough explains Pydantic AI 2.0 through a support agent built in both version 1.0 and version 2.0. The presenter describes a capability as a reusable unit containing instructions, tools, lifecycle hooks and model settings. The example separates knowledge base access from human escalation, then reuses the knowledge base capability in an FAQ agent.
The code comparison shows how capabilities organize responsibilities that the earlier example defines together. It also covers progressive disclosure: the agent receives brief capability descriptions and loads full instructions when needed. In the CLI demonstration, a question about Slack triggers knowledge base search. A duplicate billing complaint triggers escalation. The presenter explicitly says the ticket creation is mocked, so the demo does not establish a working refund or support integration.
The final section explains the framework's lean core and harness, including sandboxed code execution and a brief mention of Monty. The presenter also compares Pydantic AI with the Claude Agent SDK and Codex SDK, describing those SDKs as easier for personal agents but slower and more token intensive. These are the presenter's assessments, rather than benchmark results.
A sponsored segment introduces Nimbalyst, an open source visual workspace for a coding assistant. The presenter shows file diffs and parallel session management, and says its workspace data lives locally as Markdown files. The main tutorial does not demonstrate local model inference or offline operation.
