Agentic AI with Pydantic AI, A2A and MCP

This course builds an incident-triage agent with Pydantic AI and Amazon Nova Lite through AWS Bedrock. The agent returns structured severity, summary and rationale fields. The instructor exposes it through FastA2A and Uvicorn, then sends JSON-RPC requests and polls task IDs in Postman.

The implementation adds DynamoDB storage and SNS email notifications. After correcting missing environment variables, the instructor verifies a database record and a received email. A context-based guard limits repeated tool calls. The setup uses hosted inference and AWS resources; it does not demonstrate local model execution or offline operation.

The MCP lessons cover initialization, tool discovery and calls against a public developer-documentation endpoint. MCP Inspector displays calls and request history, while Cline retrieves documentation through a configured remote endpoint. The Pydantic AI integration also invokes the documentation tool, though the instructor does not establish the relevance of its results.

Later sections explain multiagent patterns and show a research-agent/writer-agent code example without running it. Hosted ChatGPT agent demonstrations encounter calendar and web-search failures. Ordinary chat succeeds in separate attempts, so those results do not establish a working autonomous calendar or research workflow.