Workshop: Building Durable Deep Research Agents

Using PydanticAI, Temporal, and Pydantic Logfire Simple LLM interactions work fine, but as you build more complex systems with longer-running workflows, failures become expensive. When your agent has completed several time-consuming steps (searches, data processing, analyses) and then crashes, losing all that progress creates a terrible user experience and wastes significant compute resources. Durable execution solves this by automatically saving workflow state as your agent runs. If interrupted, the system replays completed steps instantly using cached results, then continues from the exact point of failure. No restarting from scratch, no lost compute, no frustrated users. In this workshop you will create a production-ready Deep Research Agent that plays 20 Questions, but instead of a human guessing, multiple LLM agents will work together to find the answer. You will learn how to use Pydantic AI to build a toy example system with multiple agents that will structure a research plan, run in parallel to gather information, and synthesize results into conclusions. You will use Temporal agent wrappers to ensure the system can recover from failures and see into every step of execution with observability using Pydantic Logfire. You will also learn how to evaluate different model’s performance by using Pydantic AI Evals and visualise them on Pydantic Logfire. Workshop Outcomes By the end, you will: Understand when and how to implement durable execution for AI agents Have working code for a resilient multi-agent system you can adapt Know how to add observability to debug and monitor agent behavior Be able to handle failures gracefully without expensive re-computations Learn how to gauge models performance by using Evals Setup Requirements Software Installation Python 3.10 or higher (3.12+ recommended) Code editor (VS Code, PyCharm, Zed, or similar) Git (for cloning the workshop repository) Accounts to Create Pydantic Logfire - Free tier available, sign up before the workshop No LLM API keys needed! - We will provide free inference through the Pydantic AI Gateway What to Review Before the Workshop Required Knowledge Python fundamentals - async/await, type hints, decorators, context managers Pydantic Validation basics - models, validation, field types, serialization LLM concepts - prompts, function/tool calling, structured outputs Basic async patterns - task groups, concurrent execution, error handling Recommended Reading: PydanticAI documentation (especially the introduction and agents section) Temporal core concepts (workflows vs. activities) PydanticAI + Temporal integration Link to GitHub repository https://github.com/pydantic/pydantic-stack-demo/tree/main/durable-exec