Samuel Colvin

Samuel Colvin is the founder and chief executive of Pydantic and the creator of Pydantic, the open-source Python library that makes type annotations enforceable at runtime. He has extended its approach to reliable data validation into infrastructure for building, observing, evaluating, and securely running AI applications.

Before founding Pydantic’s commercial company, Colvin co-founded TutorCruncher. His other open-source projects include python-devtools and watchfiles, a Rust-backed Python file-watching library.

He founded the company behind Pydantic in February 2023 with Sequoia-led seed funding while keeping its original library open source. Pydantic V2 followed in June 2023 with a Rust-based validation core. In October 2024, Colvin launched Pydantic Logfire and announced a $12.5 million Series A. The company’s products now include Pydantic AI, an agent framework; Pydantic Logfire, an observability platform; and Pydantic Evals, an open-source evaluation framework announced in April 2025.

How Colvin thinks about dependable agents

• Type-safe AI applications: Explicit schemas, typed dependencies, and checked tool interfaces make agents easier to modify without introducing errors. When a model produces invalid structured output, Pydantic AI can return the validation failure and request a correction. Colvin also emphasizes defining when an agent loop should terminate, whether through structured output or a designated final-result tool. His demonstration of typed agents and validation-driven retries shows these mechanisms in practice.

• Observability across the whole application: Colvin considers AI observability a capability of broader software observability, not a durable standalone category. Logfire combines OpenTelemetry-based traces, logs, and metrics with visibility into model calls, tool execution, failures, latency, and cost. Sensitive deployments can restrict recorded content, retain categorical evaluation results, or operate within self-hosted infrastructure.

• MCP sampling: This Model Context Protocol feature lets specialized agentic tools request inference through their client instead of maintaining separate model access. Colvin’s BigQuery and PyPI example combines SQL validation, retries, progress reporting, and tracing while keeping tool-specific instructions out of the main agent’s context.

• Durable agent execution: Integrating Pydantic AI with Temporal allows interrupted workflows to resume by replaying completed model and tool interactions instead of paying to repeat them. His multi-agent and deep-research examples use ordinary imperative Python and parallel tasks. They also illustrate why speed and cost comparisons are meaningless when output correctness goes unchecked.

• Genetic prompt optimization: Colvin uses GEPA, reference datasets, and production traces to test candidate prompts, while typed managed variables enable changes to prompts and models without redeployment. He emphasizes the practical limitations: imperfect reference answers, overfitting, inflated prompts, model-specific tuning, and the difficulty of evaluating open-ended tasks. Optimization becomes especially useful for high-volume workloads or specialized private data, as his production optimization workshop demonstrates.

Sandboxing generated code

Pydantic Monty extends Colvin’s focus on explicit boundaries to AI-generated code. The Rust-based Python interpreter begins without filesystem, network, or environment access and grants capabilities selectively. A unified sandbox interface also supports fuller CPython environments when applications need additional packages.

Colvin tested Monty through escalating public security challenges: an initial escape was found and repaired, a second round produced no successful escape, and a third challenge offered $20,000 in August 2026. His public invitation to break the sandbox makes adversarial testing central to the product’s security claims.