PydanticAI Agents Workshop

Build, validate, test, evaluate, instrument, and hand off a typed production-minded Python agent with PydanticAI V2.

A rigorous, implementation-first workshop in which learners build a bounded professional Python application with PydanticAI V2. Learners define typed agent boundaries, inject runtime services through dependencies and RunContext, validate structured outputs, expose safe tools and toolsets, choose explicit control flow, engineer bounded failure and human-control paths, test and evaluate representative cases, inspect traces, and deliver a production-minded implementation/handoff package. The workshop complements the vendor-agnostic AI Agent Course and AI Agent Frameworks Course by concentrating on current PydanticAI APIs and related components only where they improve the selected build. A fictional service-operations scenario and synthetic records provide a no-paid-service completion path; learners may substitute approved credentials and systems. Mastery requires at least 80% overall and satisfactory completion of the capstone's critical safety, validation, testing, and handoff criteria.

What you will be able to build

• Implement a runnable Python agent using current PydanticAI V2 Agent, model, instructions, dependencies, RunContext, output_type, and run APIs.
• Define typed dependency, tool-input, tool-result, and structured-output contracts that expose validation failures rather than silently accepting malformed behavior.
• Connect at least one meaningful local, simulated, or credential-backed external action through a function tool or reusable toolset with explicit permissions and side-effect boundaries.
• Select and implement the simplest appropriate control-flow pattern among deterministic application orchestration, delegation, programmatic handoff, Pydantic Graph, capabilities, dynamic toolsets, and MCP.
• Handle tool, model, provider, timeout, validation, approval, interruption, and duplicate-side-effect risks with bounded retries and documented recovery behavior.
• Create representative automated tests and Pydantic Evals datasets/evaluators that measure successful behavior and meaningful failure cases.
• Instrument and inspect agent runs through OpenTelemetry-compatible tracing, using Logfire when available, and document debugging, latency, usage, and cost evidence without inventing benchmarks.
• Defend architecture and model/provider choices using measured evidence, model capability checks, security boundaries, and explicit version-sensitive assumptions rather than framework hype or one successful demonstration run evidence—or one successful demonstration run alone-. actually must fix. Keep valid string.