⚡️Monty: the ultrafast Python interpreter by Agents for Agents — Samuel Colvin, Pydantic

The sandbox latency bottleneck

Monty fills the critical gap between simple tool calling (safe but limited) and full sandboxes (powerful but slow), solving the 1-second+ cold start times of solutions like Daytona that hinder high-frequency agent workflows.

Enterprise self-hosting constraints

Large financial institutions cannot use cloud sandboxes like Modal or E2B due to external infrastructure requirements and compliance needs, demanding an in-process solution installable as a single binary without complex orchestration.

Sub-microsecond execution speeds

Monty achieves 800-nanosecond execution times from code to result in hot loops, compared to 1+ seconds for WebAssembly or containerized alternatives, enabling real-time agent tool calling.

WebAssembly security failures

Pyodide-based approaches require 62MB of dependencies (Deno runtime + packages) and cannot prevent memory exhaustion attacks or JavaScript escapes, whereas Monty's Rust implementation provides inherent memory-safe isolation.

Single binary deployment

Distributed as a standalone Rust binary installable via pip, npm, or soon Dart and Kotlin, Monty eliminates complex dependency chains while supporting any platform that runs Rust.

LLM-accelerated implementation

Colvin used AI to implement 20+ Python built-in functions in hours rather than weeks by leveraging the model's existing knowledge of Python internals and trivial unit testing against CPython output.

External function security model

Rather than supporting third-party packages or class definitions internally, Monty routes complex operations like HTTP requests and Pydantic validation through secure external calls to the host runtime.

Community-driven AI contributions

Contributors are using LLMs to automatically implement standard library modules, with one recent PR adding 50 math functions and 800+ tests generated entirely by AI agents.

Programmatic tool calling

Inspired by conversations with Anthropic engineers about type safety for agentic workflows, Monty specifically targets the "code mode" pattern where 70% of sandbox usage involves glorified tool calling for calculations and chart rendering.

Logfire observability integration

Pydantic's observability platform leverages Monty's approach to let AI agents write arbitrary SQL against telemetry data, providing analytical capabilities unavailable on restricted platforms like Langsmith or Braintrust.