Playground in Prod - Optimising Agents in Production Environments
AI Engineer Europe 2026 · 1:20:40
AI developer tools, data validation and observability
Pydantic builds tools for developers to validate application data, build AI agents and monitor production software. Its open-source Pydantic Validation library checks and serializes data using Python type annotations, while Pydantic AI provides an agent framework and Pydantic Evals supports code-based evaluations. Logfire brings traces, logs and metrics together through OpenTelemetry, helping teams inspect agent behavior and improve applications. Logfire AI Gateway adds model routing and spending controls.
Founder and CEO Samuel Colvin started the Python library in 2017 and established the commercial company in 2022. Its engineering approach connects Python’s type system with a Rust validation core; models can also generate JSON Schema for integration with other tools. Pydantic extended that approach to agent code execution with Monty, an open-source Python interpreter written in Rust. Monty lets agents express loops, conditionals and asynchronous calls while restricting interaction with the outside world to functions explicitly supplied by developers.
In 2024, Logfire attracted 2,000-plus developers and 150 companies during its beta phase. Its commercial observability business launched Logfire out of beta that year, alongside a $12.5 million Series A led by Sequoia, with participation from Partech and Irregular Expression.
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here to learn how workflow IDs enable resumption and how replay handles previously completed LLM calls in twenty-questions and deep-research examples.
Samuel ColvinAI Engineer Code 2025
Use the BigQuery demonstration to see how Pydantic AI queries PyPI download data with retries and observable execution.
Samuel ColvinAI Engineer World's Fair 2025
Follow the relationship-extraction example to learn how evaluation datasets and managed variables support improvements to deployed agents.
Samuel ColvinAI Engineer Europe 2026
Samuel Colvin explains why type safety matters when building and refactoring AI applications, then demonstrates Pydantic AI agent loops, structured extraction, schema validation, validation-driven retries, and typed dependencies.
Samuel ColvinAI Engineer World's Fair 2025
Schema validation and validation-driven retries address agent outputs, while Temporal activity wrapping and persisted workflow state support recovery after crashes. These sessions connect application correctness with continuity of long-running execution.
Colvin argues for MCP as a communication mechanism among autonomous agents. He contrasts it with OpenAPI and explains how sampling gives an agentic MCP server access to an LLM through its client.
The production workshop demonstrates golden-answer evaluation and GEPA-driven prompt optimization. Pareto-frontier candidate selection frames improvements alongside practical constraints such as model latency and prompt size.
Affiliations reflect each recorded session, not necessarily current employment. Despite its title, Human seeded Evals — Samuel Colvin, Pydantic does not cover evaluations.