Pydantic

Pydantic builds developer tools for Python and AI, including the Pydantic validation library, Pydantic AI for agents, Logfire for observability and Monty for executing agent-written Python. Its expansion combines validated structured outputs, typed agent workflows and application tracing. Pydantic AI V1, released in September 2025, made API stability a central commitment; in March 2026, the company announced that its AI Gateway would consolidate into Logfire, bringing model routing and observability into the same account.[7][9][10]

These AI tools followed an earlier Python project. The library’s type-hint schemas, JSON Schema generation and Rust validation core provide context for the company’s later work on structured outputs and production agents.[1][5][7]

Samuel Colvin dated Pydantic’s beginnings to 2017 in his February 2025 interview. The library uses Python type hints to define data schemas and supports validation and serialization. By default, it also coerces compatible input: a string such as "123" can become the integer 123. Strict mode disables that coercion.[1]

Colvin credited Sebastian Ramirez, the creator of FastAPI, with adding JSON Schema support before Pydantic version one to serve OpenAPI. He recalled receiving roughly 50 commit-related emails over a weekend. That contribution later allowed Pydantic to generate schemas as well as validate data, providing a shared definition for LLM structured outputs and tools. Colvin did not claim involvement in OpenAI’s choice of JSON Schema.[1]

On February 16, 2023, Colvin announced that he was starting a company around Pydantic and had closed a seed round led by Sequoia. Participants included Partech, Irregular Expressions and angels Bryan Helmig, Tristan Handy and David Cramer. The plan was to finish V2, continue maintaining the library and build commercial cloud services and developer tools while keeping Pydantic open source under the MIT license. The announcement reported 46 million PyPI downloads in the preceding 30 days as of January 31, 2023—a package-download count, rather than a count of users or customers.[3]

Pydantic announced its first V2 alpha on April 3, 2023, followed by the official release on June 30; both announcements credited Samuel Colvin and Terrence Dorsey. Validation moved into the separate Rust package pydantic-core, while the Python library largely generated core schemas defining validator and serializer behavior. The alpha announcement claimed V2 was 5–50 times faster than V1; that range described library performance, not an application-wide speedup.[5][4]

On October 1, 2024, Pydantic launched Logfire out of beta and announced $12.5 million raised in Series A funding, led by Sequoia with existing investors Partech and Irregular Expression. In February 2025, Colvin described the goal as general-purpose observability with first-class AI support, so developers could see model calls alongside the rest of their application. He distinguished the MIT-licensed Pydantic and Pydantic AI libraries from the then-closed-source Logfire platform.[6][1]

Logfire lets users query telemetry directly with SQL. Colvin explained that this gave developers room to devise their own analyses and evaluations rather than depend entirely on predefined views. He also described sending span information when a span starts, as well as when it finishes, so long-running model calls and batch jobs become visible before completion.[1]

According to Colvin’s February 2025 account, Logfire’s database moved from ClickHouse to Timescale to DataFusion. The choice reflected the importance of SQL that end users could write, open-source infrastructure and the team’s Rust expertise. He recalled implementing JSON support using Pydantic Core’s parser and improving a string-comparison kernel. DataFusion maintainer Andrew Lamb advised the company. Colvin acknowledged that the database changes slowed delivery; his criticism of ClickHouse’s JSON support concerned the period when the team tried it, and he noted that support had since improved.[1]

By April 2, 2025, Pydantic had announced Self-Hosted and Enterprise Cloud editions of Logfire, along with a Model Context Protocol server giving agents access to observability data. The same update introduced Pydantic Evals, an open-source framework for benchmarking and evaluating AI models with optional Logfire integration.[8]

Colvin said Pydantic AI grew from dissatisfaction with existing frameworks’ production readiness. Its agents combine a prompt, tools and an optional structured return type. The framework emphasizes type safety, dependency injection and unit-tested documentation examples. Its V1 release supported validated structured output through tool calling, provider-managed outputs or manual prompting, and emitted semantic-convention-compliant OpenTelemetry data for observability.[1][7]

Colvin initially resisted graphs in favor of ordinary Python control flow, but examples of complex workflows changed his view. Pydantic’s graph approach uses nodes, normally defined as dataclasses, that return the next node to execute; return-type annotations define the possible transitions and can generate Mermaid diagrams. In the February 2025 interview, he said the agent implementation had moved onto this graph library without changing its public API. Per-node database persistence was still planned at that point.[1]

Pydantic AI V1 arrived on September 4, 2025. The company reported 15 million downloads after nine months of feedback, iteration and usage. It committed to at least six months without breaking users’ code and at least six months of V1 security fixes after a future V2 release. V1 also added human approval for tools and durable execution with Temporal, which the company described as out of beta and production-ready; durable execution for graphs remained separately listed as forthcoming.[7]

Pydantic also built the open-source browser sandbox Pydantic.run to demonstrate Logfire and Pydantic AI. Colvin connected the project to onboarding friction: people signed up for Logfire but could drop off when asked to open a development environment and configure a project. He chose a terminal-like interface over Marimo to avoid suggesting that Logfire was only for notebooks. Runnable documentation examples supplied code for the sandbox; model-proxy support was still planned in February 2025.[1]

On February 27, 2026, Pydantic announced Monty, an open-source Python interpreter written in Rust for agent-generated code. Its bytecode VM uses Ruff’s parser, bundles ty for type checking and supports snapshots that serialize execution state for later resumption. Outside interaction occurs through explicitly supplied external functions. The default grants no filesystem, network or environment-variable access, with memory, recursion and execution-time limits. At announcement, Monty supported a subset of Python and did not support arbitrary third-party packages.[9]

Monty lets a model express loops, conditionals and parallel calls in Python rather than request each tool step separately. In Pydantic’s illustrative London-versus-Paris weather demo using Claude Sonnet 4.5 and demo endpoints, sequential tool calling required four LLM round trips; a generated script coordinating the same host functions required two. This was a publisher demonstration, and the Pydantic AI CodeExecutionToolset integration was described as forthcoming.[9]

On March 23, 2026, Pydantic announced consolidation of its AI Gateway into Logfire. Colvin cited the friction of separate accounts, API keys, dashboards and billing relationships for routing model calls and observing them. The combined product correlated gateway requests with application traces and included a traced prompt-testing playground. Consolidated invoicing was still described as work in progress.[10]