Descriptions:
Latent Space sits down with Samuel Colvin, creator of Pydantic and Logfire, for a deep-dive into Monty—a new ultrafast Python interpreter written in Rust and designed specifically to serve AI agents. Colvin traces the project’s origin to a series of independent conversations with Anthropic engineers, each of whom privately emphasized that type safety is critical when chaining LLM tool calls. That recurring signal, combined with Anthropic’s December release of programmatic tool-calling and Cloudflare’s popularization of ‘code mode,’ convinced Colvin to revisit an earlier abandoned project and build Monty out properly over Christmas—producing roughly 30,000 lines of Rust.
Monty is designed to sit between simple function-calling (safe, low overhead, but limited expressiveness) and full sandbox environments (powerful but costly and complex to operate). Colvin explains that an investor surveying the sandboxing space estimated that around 70% of sandbox invocations are essentially glorified tool calls—chart rendering, calculations, data transformations—tasks that don’t need full computer-use capabilities but where executable code is far more powerful than JSON schemas. Monty targets exactly that middle tier, with type safety as its core design guarantee.
The conversation also covers Pydantic Logfire, the team’s AI-native observability platform competing with Braintrust and LangSmith, and how the Pydantic team uses Devin alongside a custom-trained internal review tool for pull request workflows. Colvin shares his personal coding stack—primarily Claude Code, supplemented by Gemini CLI—and discusses how AI-generated code is reshaping what a small team can realistically build and maintain.
📺 Source: Latent Space · Published March 14, 2026
🏷️ Format: Interview
